University of Southern California Dissertations and Theses

Magnetic induction-based wireless body area network and its application toward human motion tracking

(USC Thesis Other)

Magnetic induction-based wireless body area network and its application toward human motion tracking

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Magnetic Induction-Based Wireless Body Area Network and Its Application Toward Human Motion Tracking by Negar Golestani A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (ELECTRICAL ENGINEERING) May 2021 Copyright 2021 Negar Golestani To my family and In loving memory of my grandfather. ii Acknowledgements First of all, I would like to express my sincere gratitude to my advisor and committee chair, Prof. Mahta Moghaddam, for the continuous support, motivation, and encouragement throughout my graduate studies. Her patience, sincerity, and immense knowledge have deeply inspired me. It was a great privilege and honor to work and study under her guidance. I would also like to thank the rest of my committee members: Prof. Gianluca Lazzi, Prof. Constantine Sideris, and Prof. Greg Ver Steeg, for their time, suggestions, and insightful comments. I would like to thank my fellow labmates in the Microwave and Systems, Sensors and Imaging Laboratory (MiXIL) for the fun-time, discussions, and friendship over the years. I would also like to thank my close friends and family for always being there through ups and downs and giving me so much love and support over the years. A special thanks to my family. Their unconditional love, support, and kindness have always brightened my day, even while living thousands of miles away. I would like to express my deepest appreciation to my parents, whose encouragement and sacrifices at every step of my life have helped me strive towards my goals. Without them, I would certainly not have reached this point, and I am forever grateful to them. I also dedicate this thesis to my parents and the memory of my grandfather, whose role in my life was, and remains, immense. iii Table of Contents Dedication ii Acknowledgements iii List of Tables vi List of Figures viii Abbreviations xii Abstract xiii 1 Introduction 1 1.1 Background and Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.4 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2 Magnetic Induction (MI) Communication 11 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.2 Theoretical Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.2.1 Mutual Inductance of Two Arbitrarily Oriented Coils . . . . . . . . . . . . 13 2.2.2 Two-port Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.2.3 Lumped Reactive Impedance Matching . . . . . . . . . . . . . . . . . . . 21 2.2.4 Power Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.3 Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.3.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.3.2 Human Body Model and Dielectric Properties . . . . . . . . . . . . . . . . 27 2.3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.4 Model Accuracy Analysis in Generating Time-series Data . . . . . . . . . . . . . 34 2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 3 Human Activity Recognition 39 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.2 System Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.3 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 iv 3.4 Synthetic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.4.1 Magnetic Induction System Setup . . . . . . . . . . . . . . . . . . . . . . 44 3.4.2 Motion Capture Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.4.2.1 Biological Motion Library (BML) . . . . . . . . . . . . . . . . 46 3.4.2.2 Multimodal Human Action Database (MHAD) . . . . . . . . . . 47 3.4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.5 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.5.1 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.5.2 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.5.2.1 Bag of Words (BoW) . . . . . . . . . . . . . . . . . . . . . . . 52 3.5.2.2 Machine Learning-based Classification . . . . . . . . . . . . . . 54 3.5.2.3 Recurrent Neural Network (RNN) . . . . . . . . . . . . . . . . 61 3.5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4 3D Motion Tracking 71 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 4.2 System Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 4.2.1 Operating Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 4.2.2 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 4.2.3 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 4.3 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 4.3.1 Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 4.3.2 Synthetic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 4.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 4.4.1 Machine Learning-based Regression . . . . . . . . . . . . . . . . . . . . . 83 4.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 4.5 Magnetic Induction vs. RFID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 4.5.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 4.5.2 Measurement Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 4.5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 5 Conclusion and Future Work 99 Bibliography 103 v List of Tables 2.1 The magnetic induction (MI) coils parameters used for model validation. . . . . . . 26 3.1 The marker pairs defining two ends of a body part. . . . . . . . . . . . . . . . . . 45 3.2 Hyperparameters related to the deep recurrent neural network architecture and learning and their range of values explored for tuning during training. . . . . . . . 64 3.3 The performance summary of classification results using generated synthetic mag- netic induction (MI) motion data of different datasets. . . . . . . . . . . . . . . . . 66 3.4 Performance comparison with other state-of-the-art methods using different modal- ities for activity detection. Multi-Task Conditional Restricted Boltzmann Ma- chines (MT-CRBMs). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.1 Performance of regression models in motion tracking using synthetic data gener- ated for the different settings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 4.2 Performance of regression models in motion tracking using RSSI data measured by the RFID reader. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 vi List of Figures 2.1 Magnetic induction-based wireless body area network (MI-WBAN). . . . . . . . . 12 2.2 Maximum operating frequency that satisfies small loop antenna assumption as a function of coil radius. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.3 Acceptable operating frequency as a function of maximum distance between coils (range). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.4 Relative location and alignment of the transmitter (TX) and receiver (RX) coils in different coordinate systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.5 Equivalent Two-port network model of magnetic induction (MI) system. . . . . . . 20 2.6 Lumped reactive matching networks. . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.7 HFSS simulation setup for model validation. . . . . . . . . . . . . . . . . . . . . . 25 2.8 Measurement setup for model validation. . . . . . . . . . . . . . . . . . . . . . . . 25 2.9 Magnetic induction (MI) node prototype. . . . . . . . . . . . . . . . . . . . . . . 26 2.10 Measurement setup with matching circuit for model validation. . . . . . . . . . . . 26 2.11 Dielectric properties of human body tissues. . . . . . . . . . . . . . . . . . . . . . 27 2.12 The power gain of two perfectly aligned circular coils separated by relative dis- tance of d in the background medium air. . . . . . . . . . . . . . . . . . . . . . . 29 2.13 The power gain of two circular coils separated by distance of 40 cm, and surface misalignment ofq in the background medium air. The location of transmitter (TX) and receiver (RX) coils are fixed and angular misalignment (q) varies between 0 to 180 degrees. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.14 Simulation and experimental measurement results. The power gain comparison of magnetic induction (MI) system with and without matching networks. . . . . . . . 31 2.15 The NRMSE of estimated forward voltage gain with respect to simulation and experimental measurements vs. frequency (left), andq x (right). . . . . . . . . . . . 33 2.16 Schematic representation of measurement setup. . . . . . . . . . . . . . . . . . . . 35 2.17 A camera frame sample after processing. . . . . . . . . . . . . . . . . . . . . . . . 36 2.18 Measured vs. synthetic magnetic induction (MI) data. The measured and simulated voltage gain of two MI coils during arbitrary movements, such that both relative alignment and location of coils vary. . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.1 Magnetic induction-based human activity recognition (MI-HAR) system. . . . . . 41 3.2 Location of MI transceivers on the human body. . . . . . . . . . . . . . . . . . . . 44 3.3 Location of markers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 3.4 The center and alignment of a bone and its corresponding coil that can be calcu- lated using markers locations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 vii 3.5 Synthetic magnetic induction (MI) motion data. The forward voltage gain S 21 between the receiver (RX) and transmitters (TX 1 -TX 8 ) are generated using the proposed MI model and the human motion data captured for different activities in Biological Motion Library (BML). . . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.6 Synthetic magnetic induction (MI) motion data. The forward voltage gain S 21 between the receiver (RX) and transmitters (TX 1 -TX 8 ) are generated using the proposed MI model and the human motion data captured for different activities in Berkeley Multimodal Human Action Database (MHAD). . . . . . . . . . . . . . . 49 3.7 Synthetic magnetic induction (MI) motion data. The forward voltage gain S 21 between the receiver (RX) and transmitters (TX 1 -TX 8 ) are generated using the proposed MI model and the human motion data captured for different activities in Berkeley Multimodal Human Action Database (MHAD). . . . . . . . . . . . . . . 50 3.8 The process of calculating the bag-of-words representation of time-series motion data samples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.9 Architecture of deep recurrent neural network (RNN). . . . . . . . . . . . . . . . . 63 3.10 Average R-squared between XYZ of each target point and data of its corresponding accelerometer and magnetic induction (MI) transceiver. . . . . . . . . . . . . . . . 65 3.11 Confusion matrix for the validation set corresponding to the Biological Motion Library (BML). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 3.12 Confusion matrix for the validation set corresponding to the Berkeley Multimodal Human Action Database (MHAD). . . . . . . . . . . . . . . . . . . . . . . . . . . 68 4.1 Data-driven location tracking across all settings for a predefined target point. . . . . 75 4.2 The MI sensor prototype hardware contains 1) variable capacitor for frequency tuning, 2) an envelope detector, and 3) Arduino microcontroller for measurement. . 77 4.3 Schematic representation of measurement setup. . . . . . . . . . . . . . . . . . . . 79 4.4 The measured and simulated induced voltage at an MI sensor during two arbitrary movements, such that both relative alignment and location of the coil varies. . . . . 82 4.5 The evaluation process, including training with synthetic data, evaluation on syn- thetic data, and test on the real-word measurements. . . . . . . . . . . . . . . . . . 86 4.6 Tracking performance metrics across all configuration settings on the measured motion and MI data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 4.7 Distance and motion tracking in the z-direction. . . . . . . . . . . . . . . . . . . . 89 4.8 The designed RFID tag with customized antenna and variable capacitor. . . . . . . 93 4.9 An experimental RFID measurement including RSSI and motion data. . . . . . . . 94 4.10 Comparison of MI signal and RSSI of passive RFID in 3D motion tracking using statistical measures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 viii Abbreviations ADLs activities of daily living AID automated insulin delivery ANN artificial neural network AoA angle of arrival A WG American wire gauge bagging bootstrap aggregating BLE Bluetooth low energy BML Biological Motion Library BoW bag-of-words CART classification and regression trees DBN deep belief network DC direct current DRNN deep recurrent neural network DT decision trees DWT discrete Wavelet transform EEPROM electrically erasable programmable memory ELBO evidence lower bound EM electromagnetic ET extremely randomized trees or extra trees FEM finite-element method FFT fast Fourier transform FTM fast transfer mode ix HAR human activity recognition HBC human body communication HCI human-computer interface HF high frequency HFSS high frequency structural simulator IC integrated circuit ICs integrated circuits IEEE Institute of Electrical and Electronics Engineers IMU inertial measurement unit IMS industrial, scientific and medical KL Kullback-Leibler KNN K-nearest neighbors KPCA kernel principal component analysis LDA linear discriminant analysis LF low frequency LightGBM light gradient boosting machine LR linear regression LR logistic regression LSTM long short-term memory MAC medium access control MAE mean absolute error MAPE mean absolute percentage error MCODE molecular complex detection MHAD Multimodal Human Action Database mHBC Magnetic Human Body Communication MI magnetic induction MI-HAR magnetic induction-based human activity recognition x MI-WBAN magnetic induction-based wireless body area network MLP multi-layer perceptrons MMH maximum marginal hyperplane MoCap motion capture MSE mean-squared error NB narrowband NFC near-field communication NRMSE normalized root-mean-squared error PHY physical PoA phase of arrival QoS quality of service R 2 R-squared RAD radiation RBF radial basis function ReLU rectified linear unit RF random forest RFID radio frequency identification RMSE root-mean-squared error RNN recurrent neural network RSSI received signal strength indicator RX receiver SAR specific absorption ratio SDK software development kit SVM support vector machines TDM time-division multiplexing ToA time of arrival TX transmitter xi UHF ultra-high frequency UID unique identifier UWB ultra-wideband V AE Variational Auto-Encoder VNA vector network analyzer VR virtual reality WBAN wireless body area network WHO World Health Organization WPAN wireless personal area network WSN wireless sensor network xii Abstract A wireless body area network (WBAN) is a wireless network of wearable or implanted sensors with different medical and non-medical applications. One of these applications is in human ac- tivity recognition. Recognizing human physical activities using sensor networks has attracted sig- nificant research interest in pervasive computing due to its broad range of applications, such as rehabilitation, athletics, and senior monitoring. However, there are critical challenges inherent in the sensor-based activity recognition method, such as wireless communication technology, power consumption, coverage, and reliability. Magnetic induction (MI) is a new physical layer technique, which can address some of these problems associated with wireless communication. Moreover, a proper classification method can tackle the challenges and constraints associated with the detection stage. We begin by investigating the development of the magnetic induction technique as a new phys- ical layer for sensor networks working around the human body. To facilitate performance analysis of the MI wireless system around the human body, we introduce a circuit model describing power transmission between the transceivers. This model exploits an analytical representation of the mutual inductance between two coils defined as a function of operating frequency, and coils’ ge- ometry, relative position, and orientation. It provides an accurate and fast method of calculating mutual induction, a key parameter in the optimal design of an MI communication system. We then verify the model via experimental measurements and numerical simulations and investigate the effect of impedance matching on system efficiency. The validation results are encouraging and xiii indicate that for frequencies up to 30 MHz, the human body does not affect the performance of MI communication system. We also explore the capability of the model in synthesizing time-series data using experimental measurements and comparison with ground truth data. Next, we explore the application of the MI system in human activity recognition. Critical chal- lenges are inherent in designing a wireless network operating around a lossy medium such as the human body to gain a trade-off between power consumption, cost, computational complexity, and accuracy are described. We propose an innovative wireless system based on magnetic induction for human activity recognition to tackle these challenges and constraints. The proposed frame- work consists of two stages: data acquisition and detection. For the first stage, we employ the MI system model to synthesize MI motion data corresponding to several physical activities. In the second stage, machine learning techniques and deep recurrent neural networks are applied to the synthetic data to classify human movements. We present that MI signals are informative de- scriptors for the motion of human body parts, and therefore this approach can successfully identify human movements. Finally, we extend the magnetic induction-based human activity recognition (MI-HAR) system and study the application of MI system in motion tracking instead of identifying the user’s activity. Analyzing body movement in 3D space facilitates behavior recognition in different applications, including healthcare, rehabilitation, sports, and senior monitoring. We propose a motion tracking system based on MI to address issues and limitations of existing motion tracking systems based on wearable sensors. We integrate a realistic prototype of an MI sensor with machine learning techniques and investigate one-sensor and two-sensors configuration setups for motion reconstruc- tion. The approach is then evaluated using measurements and synthesized datasets generated by the analytical model of MI system and laboratory measurements. We show that the system has an acceptable dynamic spatial resolution compared to the ground truth variation captured by Kinect. xiv Chapter 1 Introduction 1.1 Background and Motivations Recent advances in sensor design and miniaturization combined with low-power wireless com- munications and data analysis allow for the design and development of wearable device networks operating around the human body with wireless communication capabilities. Wearable devices have enabled the feasibility of continuous monitoring of users’ health markers such as heartbeat, blood pressure, oxygen saturation, body temperature, blood glucose [1], which had made them a promising technology with different applications in medical and non-medical fields [2–4]. It allows, for example, remote monitoring of elderly people [5] in their home [6]. According to the U.S. Bureau of the Census, the United States population aged 65 and over will experience con- siderable growth from 43.1 million in 2012 to 83.7 million in 2050 [7]. Between 1990 and 2025, the older population continues to grow at an unprecedented rate from 357 million to 761 million worldwide [7, 8]. WBAN also can be used for diagnosing and automated insulin delivery (AID) in diabetic patients [9, 10]. The World Health Organization (WHO) has reported that between 1980 and 2014, the number of people with diabetes has risen from 108 million to 422 million. Diabetes is a major cause of blindness, kidney failure, heart attacks, stroke, and lower limb amputation, and it will be the 7th leading cause of death in 2030 [11, 12]. These statistics show that providing a cost-effective health care solution will be a critical issue in the near future. 1 Many research studies have reported using smart devices, such as smartphones, smartwatches, and fitness bands, to provide useful insights about an individual’s performance and health status. These wearables have a wide variety of embedded sensors that enable multi-modal sensing and data fusion, resulting in improved accuracy and more specific inferences than single-sensor systems. They have demonstrated satisfactory performance in, for example, human activity monitoring. Re- search work [13] introduces an unsupervised classification method called molecular complex de- tection (MCODE) for physical human activity recognition using smartphone accelerometers. The experimental results and performance analysis are also presented on real-world datasets, includ- ing a public activities of daily living (ADLs) dataset and two datasets containing sports activities such as race walking and basketball. The results show that the method is practical for recognizing real-world physical activities and is more effective than the K-means method. The work presented in [14] also proposes a real-time activity monitoring algorithm using sen- sors data from a smartphone, a wristband, or their combination. The algorithm identifies devices available on the body and uses an interval of walking to detect the gravity and normalize the de- vice’s orientation. The normalized data is then employed to find the smartphone position and select a location-specific classification model for activity detection. This work has evaluated algorithm steps, device configurations, and decision fusion benefits in addition to the average accuracy re- ported for activity recognition. [15] presents a smartphone inertial sensors-based approach for human activity recognition. In this work, efficient features, such as mean, median, autoregres- sive coefficients, etc., are first extracted from the raw data. Then for robustness, they are pro- cessed by a kernel principal component analysis (KPCA) and linear discriminant analysis (LDA). The processed features are then used for training a deep belief network (DBN) model for activity recognition. The proposed approach is evaluated, and the reported results are compared with con- ventional expression recognition approaches such as multi-class support vector machines (SVM) and artificial neural network (ANN). 2 Another work, [16] surveys recent advance of deep learning-based sensor-based activity recog- nition. It summarizes the literature and highlights the research work in three main categories: sensor modality, deep model, and application. Furthermore, grand challenges and feasible solu- tions are presented and discussed in detail. Moreover, [17] compares smartwatch and smartphone- based activity recognition systems. The results show that the smartwatches can identify specific hand-based activities (e.g., eating or drinking) that are not recognizable by a smartphone. Hence, a smartwatch-based activity recognition system for biomedical and health purposes is proposed that can be employed for applications such as tracking a user’s eating habits. Although Wearable devices provide a privacy-aware solution that overcomes many of the disadvantages of external, off-body monitoring approaches, they are still unable to address the specific requirements of a diverse range of applications. Single node wearable devices are limited to a single location, typically the wearer’s wrist. However, optimizing sensor positions over multiple parts of the body can significantly improve the accuracy and robustness of the system for monitoring vital signs (e.g., body temperature and heart rate) [18]. Another issue is that a single node cannot obtain adequate global information from different locations on the user’s body. For example, in human activity monitoring systems, inertial sensors embedded in a smartwatch cannot capture the movement of the legs, which restricts the system’s capability in classifying activities [19, 20]. Moreover, variations in position in systems relying on data from a single device can significantly affect the system performance [19, 21]. The work presented in [22] aims to provide a multi-sensor solution to overcome challenges associated with approaches using a single accelerometer sensor. It proposes a multi-sensor system that can achieve high recognition accuracy with lightweight signal processing algorithms running on a distributed computing-based sensor system. This work shows that a network of accelerometer sensors distributed on the body does not require complex classifiers and feature sets to achieve high recognition accuracy. In contrast, single node devices require advanced feature extraction al- gorithms and complex classifiers, which exceeds the computing ability of most existing wearable sensor platforms, to improve their overall performance. Furthermore, the authors discuss several 3 design aspects, including sampling rate, signal processing algorithms, feature selection, and clas- sifier selection, which must be considered to be able to create such a multi-sensor system. Possible solutions to achieve a trade-off between accuracy and computation time are investigated, and the comparison of six different single or multiple sensors systems is also presented. [23] studies the effect of sensor location on the user’s body and proposes a robust position-aware wearable device for a real-world scenario in which the user decides the on-body position of the device. It investi- gates the feasibility of recognizing the wearable device position using a single accelerometer and studies the influence of its on-body location in activity recognition accuracy. Due to the broad range of requirements imposed by the applications and unique characteristics of the communication channel (human body), WBAN development is challenging [22, 24] since many constraining and often conflicting requirements have to considered [25–27]. For example, the system must be inexpensive, accessible to the general public, and meet ergonomic constraints and health requirements. It must operate under proper guidelines limiting maximum power expo- sure to the user to prevent elevated temperatures in biological tissues due to specific absorption. Thus, to ensure user safety, such a system must satisfy specific absorption ratio (SAR) constraints while simultaneously providing a reliable wireless communication link [28]. Moreover, the system should guarantee the security and privacy of the user’s data at all times. Wearable devices must be small and lightweight, which places restrictions on battery size and longevity. On the other hand, frequent battery recharging may not be practical for sensor networks with multiple sensors in applications such as senior monitoring [29]. Due to the limitation of energy resources available, power management is a critical consideration during the design of these systems. Overcoming the technical challenges involved is crucial to the successful growth of wearable sensors toward widespread commercial realization [30]. Wireless communication consumes a considerable portion of the energy budget [31], and nu- merous studies have proposed and investigated low-power solutions [32–35]. Among those Blue- tooth low energy (BLE) and standards proposed by Institute of Electrical and Electronics Engineers (IEEE) such as 802.15.4 and IEEE 802.15.6 are considered as the main communication standard 4 solutions [32]. The BLE is a low power version of conventional Bluetooth technology, which was initially designed for free-space long-range (>10m) communication, rather than for a short-range WBAN. IEEE 802.15.4 wireless technology is a short-range communication system developed to provide the specifications of physical (PHY) and medium access control (MAC) layers for a wireless personal area network (WPAN) with relaxed throughput and latency requirements. This wireless technology is low power, low rate, and its main application is the implementation of WBANs. ZigBee protocol is an available option for defining the upper layers. IEEE 802.15.6 is the first international wireless body area network standard that supports communications within the surrounding area of the human body for various applications [32]. This standard defines MAC and PHY layers, where PHY schemes include narrowband (NB), ultra-wideband (UWB), and hu- man body communication (HBC). The NB and UWB technologies are RF-based techniques, while HBC is a non-RF technique that uses the human body as the communication channel. Many studies have proposed WBAN systems based on different communication protocols for various applications. For example, [36] has presented an energy-efficient WBAN system for healthcare monitoring based on ZigBee. To obtain the adaptive duty cycle, this system employs the ZigBee standard working in the beacon-enable mode. It removes the idle-mode power con- sumption and reduces the total energy consumption of sensors, extending network lifetime. The authors have shown that the proposed method be energy-efficient in long-term applications while preserving the quality of service (QoS) in terms of network performances, including an average end to end delay and network throughput. Most of these conventional state-of-the-art wireless communication technologies adopt radio wave propagation for signal transmission. Although these technologies have been successfully used in many WBAN applications, they encounter several problems [37]. For example, they are susceptible to the environment’s characteristics and experience significant attenuation around lossy media, such as the human body, which results in higher power consumption, shorter battery life, and lower reliability [34, 38, 39]. Radio-wave propagation methods are prone to interference with adjacent communication links since many devices and standards operate in the busy 2.4 GHz, the 5 industrial, scientific and medical (IMS) band [40, 41]. These far-field methods also may suffer from multi-path effects and security issues since the transmitted signal propagates a long distance through free-space, allowing far-away eavesdroppers to intercept the transmitted data. [42–45]. An alternative promising physical layer for WBAN applications is MI, which can address sev- eral issues with electromagnetic (EM)-wave propagation technologies [46]. Near-field MI has been reported as a physical layer option for near field communication systems in various applications such as hearing aid, payment cards, implants, and many more [47–49]. Although several models have been proposed for MI system, none has considered the effect of misalignment between coils. Moreover, most of them are approximate for near range communica- tions, and coils are assumed to be strongly coupled [50]. The maximum communication range for near-field wireless communication using magnetic induction is about 20 cm, which does not apply to wireless communication all over the human body [51]. Finding the best operating frequency for MI communication, where the induction regime ends and propagation effects take over, is a question that becomes relevant in the design and implementation of this system. The answer de- pends on many different factors such as coil size, number of turns in the coil, relative distance and orientation, and medium properties. In centralized sensor networks, the sensed data streams are collected from the data acquisition modalities around the user, transmitted remotely, and processed with limited resources in terms of energy availability, computational power, and storage capacity. Hence, another question that should be answered is how to process data and fuse proper signals to achieve the system objec- tive. Machine learning methods, including regression, classification, and clustering models, have been extensively used for processing sensor data. However, conventional machine learning ap- proaches rely on heuristic handcrafted feature extraction, limited by human experience or domain knowledge, which undermines the model’s performance in terms of accuracy and generalization [16, 42]. 6 Deep neural networks are another approach capable of processing more complex input data offering a higher representational power at the cost of increased computational complexity, energy, and time for training. Thus, the trade-offs between the algorithm’s computational requirements and the predictive model’s accuracy should be understood and considered. Model training is an- other concern since machine learning algorithms, specifically deep neural networks, require large datasets to train and achieve the intended generalization capabilities. Therefore, providing suffi- cient data for model training and evaluation is another critical aspect of data analysis that should be addressed. Research studies have investigated machine learning techniques and deep neural networks for wearable sensors and their application in different domains. Their comparison and statistical re- ports help to select a suitable technique for a targeted application. Studies presented in [52] and [53], for instance, demonstrate various machine learning methods for wireless sensor network (WSN)s, in addition to their advantages, disadvantages, and important parameters affecting the network and its performance. They also review algorithms for energy harvesting, synchronization, congestion control, localization, anomaly detection, fault nodes detection, routing, data aggre- gation, MAC protocols, and mobile sink path scheduling. The review presented in [54] focuses mainly on human activity recognition application. It aims to describe challenges existing in the feature extraction stage of a mobile or wearable sensor-based human activity recognition pipeline. This review suggests neural networks as a potential solution to this issue and provides a detailed summary of deep learning methods for sensor-based human activity recognition. It investigates generative, discriminative, and hybrid methods and highlights their advantages and limitations. It also discusses classification and evaluation methods and presents publicly available datasets for mobile sensor human activity detection. 7 1.2 Objectives We aim to explore and address challenges inherent in designing this wearable sensor network to understand and optimize trade-offs among power consumption, cost, computational complexity, security, and accuracy. We intend to explore the potential and benefit of using mid-range MI-based wireless communication for WBAN that allows transmission around the human body for ranges greater than 20 cm. To simplify the design and analysis of the mid-range MI-based wireless com- munication system, a model should be developed that considers various factors, such as relative location and alignment of the MI coils. The proposed model needs to be validated via simulation and experimental measurements in the presence of the human body. Then based on analysis, the efficiency of the magnetic induction-based wireless body area network (MI-WBAN) system can be enhanced by using impedance matching networks. One of the notable advantages of the pro- posed system is that the MI-based signal transmission approach opens the door to MI motion data collection. The next objective is to investigate the capability of MI system in obtaining information regard- ing the user’s body motion and physical activity. Introducing new sensing modality and sensor net- work infrastructure demand new processing algorithms for signal processing, integration, feature extraction. Development, implementation, and validation of machine learning models for achiev- ing optimum learning and decision-making are other objectives of this dissertation. The system performance, which depends on various factors, such as input features and detection algorithm, should be reported using a measure (e.g., accuracy) to prove that the proposed system is suitable for motion tracking and activity recognition. It should be shown that the MI signals received from transmitters attached around the human body represent the motions of body parts during physical activity. It means that the MI signals can be utilized as input features for the classifier. Then, a framework for detecting and recognizing human activities at the detection stage is required. Once the proof-of-concept is provided, we aim to build a realistic prototype of the system for capturing MI motion signals. Such a system would allow us to perform experiments on real-world MI data 8 to demonstrate the accuracy of movement tracking using MI system and study the environmental effects and interferences. We measure the received MI data transmitted between two MI nodes while their relative distance and alignment changes over time. The proposed system is integrated with regression models to reconstruct motion in 3D space utilizing the collected MI data. 1.3 Contributions In summary, the contributions of this dissertation include: Investigation of magnetic induction technique as a new physical layer for sensor networks working around the human body. Development and validation of a circuit model for performance analysis of an MI system with/without impedance matching networks and with/without the presence of the human body. Model accuracy analysis in generating time-series MI data. Exploration of MI system ability in human activity recognition, and providing a proof-of- concept for the proposed MI-HAR system by experimental measurements and simulations. Implementation and evaluation of machine learning-based classifiers and a deep learning framework for human activity recognition using MI motion signals and comparing with other previously introduced methods using different modalities for activity detection. Investigation of MI system ability in 3D motion tracking, and system performance assess- ment using different sensor configurations. Implementation and evaluation of machine learning-based regression models for motion es- timation using MI data. 9 1.4 Overview This dissertation consists of fives chapters. Chapter 2 explains the principle of inductive coupling and presents the circuit model that exploits the mutual inductance model between two misaligned coils. We then validate the model using experimental simulations and measurements with and without the presence of the human body. Moreover, the impedance matching technique for en- hancing the efficiency of the MI system is discussed. Then, the model accuracy in generating synthetic time-series MI data is evaluated. Chapter 3 explores the application of MI system in human activity recognition. The proposed human activity recognition system using magnetic induction-based motion signals is investigated by experimental measurements and simulations. The process of generating synthetic MI motion data using motion capture data of two publicly available datasets is presented. Several machine learning algorithms and deep learning frameworks suitable for classifying MI motion signals are presented and evaluated for the detection stage. Chapter 4 investigates the capability of MI system in tracking motion in 3D space. The MI- based sensor network is integrated with machine learning classifiers to reconstruct motion using measured MI signals. Using the analytical model of MI system, synthetic datasets are generated and utilized to train classification models. Different sensor configurations are considered, and the best approach is evaluated on real-world measurements. The process of generating synthetic data and model training and comparing several classification models are presented in this chapter. 10 Chapter 2 Magnetic Induction (MI) Communication 2.1 Introduction Magnetic induction is an emerging physical layer technique for sensor networks working around the human body. The wireless signal transmission is accomplished using inductive coupling be- tween the transmitter (TX) and the receiver (RX) instead of radiating as in traditional electromag- netic wave communication systems. The main component of each MI node is a coil, shown in Figure 2.1, because the small radiation resistance of the coil makes it an ideal candidate as a non- radiating antenna to transfer a magnetic field in the near range. Suppose the signal in the TX coil is a sinusoidal current. This current can induce a sinusoidal current in the RX coil, which enables the communication. The nodes are lightweight, portable, inexpensive, simple, and can be worn as accessories such as belts, wristbands, and jewelry [34, 55]. The MI coils have a small radiation resistance, which means that the energy propagated to the far-field is negligible. As a result, multipath fading is not an issue, and the MI system can offer a much better QoS compared to Bluetooth-type systems [34, 56, 57]. The non-propagating magnetic field produced by the coils falls off proportional to r 3 instead of r 1 for radiating fields at a transmission distance r. Although the rapid decay limits the coverage range, it can be favorable in short-range applications such as WBANs [58]. It allows the signal to remain in a ‘bubble’ around 11 Figure 2.1: Magnetic induction-based wireless body area network (MI-WBAN). the coil, which provides a personalized space for the user [44, 45]. It also minimizes the leakage outside the targeted coverage range, increases security, enables bandwidth reuse, reduces power consumption and interference [56, 59, 60]. One of the main notable advantages of the MI system is that it works well in the non-magnetic lossy environment, such as underwater, underground, and around the human body, as long as the MI system operates in the induction regime [61]. In these environments, the MI system experiences much less energy absorption than conventional radio-wave propagation technologies, as magnetic permeability of the channel medium is the most important factor affecting the MI wave [62]. It results in lower SAR for applications working around the human body. Due to smaller path loss, the MI system can transmit a signal with much less power for the same range. The MI system can be up to six times more efficient in terms of battery power than other short-range communication systems (e.g., Bluetooth) [60]. This characteristic enables many novel and demanding applications in harsh environments such as underwater monitoring of scuba divers [43, 62, 63]. This chapter is organized as follows: in Section 2.2, the principle of inductive coupling is explained, and a circuit model for the MI system is developed. The resonance technique for MI system is also explained in this section. Simulation and experiment results are presented in Section 12 2.3, including validation with a numerical method. The evaluation of the model in generating time- series data is also presented in Section 2.4. Section 2.5 discusses the results and concludes the chapter. 2.2 Theoretical Modeling 2.2.1 Mutual Inductance of Two Arbitrarily Oriented Coils A current flowing through a conducting coil generates a magnetic field, and therefore a magnetic flux. Assume that the transmitter loop is electrically small (C< 0:1l, where, C and l denote the circumference of the loop, and the wavelength, respectively) [64], and centered at the origin with the surface at xy-plane. Considering a linear, homogeneous, and isotropic background medium, one can derive the components of the magnetic field generated by TX coil (H T X ) as [65, 66]: H r = N T X S T X I T X cosq 2p g 3 ( 1 g 3 r 3 + 1 g 2 r 2 )e gr (2.1) H q = N T X S T X I T X sinq 4p g 3 ( 1 g 3 r 3 + 1 g 2 r 2 + 1 gr )e gr (2.2) H f = 0 (2.3) whereg is the complex propagation constant of the background medium (air) that can be calculated by: g = p jwm(s e + jwe 0 )=a+ jb (2.4) wherea is the attenuation constant,b is the phase constant, m is the permeability of the medium, e 0 is the real part of the complex permittivity of background medium (e), r is the distance between origin and observation point,w is the angular frequency (w = 2p f ), f is the operating frequency, N T X ;S T X ;I T X are the number of turns, area, and current that flows through the transmitter (TX) 13 Figure 2.2: Maximum operating frequency that satisfies small loop antenna assumption as a function of coil radius. Figure 2.3: Acceptable operating frequency as a function of maximum distance between coils (range). coil, respectively. Parameters e is the total conductivity defined as follows: s e =we 0 tand e (2.5) whered e is the effective loss tangent of the background medium. It is important to highlight that antennas used for WBAN applications should be small in size and wearable. The coil antennas can be used as wristbands, necklaces, bracelets, or armbands, and therefore coils up to 10 cm radius are practical for the MI-WBAN communication system. As Figure 2.2 displays, coils smaller than 10 cm radius can be modeled as a small loop antenna for frequencies lower than 50 MHz. When two coils are located nearby, some portion of generated magnetic flux will be captured by the RX coil. The ratio of total captured flux linkage in the RX coil to the current of TX coil is defined as their mutual inductance M. Reciprocity theorem states that the flux linking the primary coil, due to the current flowing through the secondary coil, is equal to the flux linking the secondary coil when the same current flows through the primary coil. The mutual inductance is defined in the standard way as (e.g., [67]): M= N RX F T R I T X = N T X F RT I RX (2.6) 14 where N RX ;I RX are the number of turns, and the current passing through the receiver (RX) coil, respectively. The quantity F RT is the magnetic flux due to I RX captured by the TX coil, and F T R is the magnetic flux due to I T X captured by the RX coil, which is defined as: F T R = Z S RX B T X :dS RX (2.7) where S RX is the area of the receiver(RX) coil, and B T X = m H T X is the magnetic flux density generated by the transmitter (TX) coil. The mutual inductance of two coils can be calculated by using (2.6) and the real parts of (2.1) and (2.2) as follows: M= m N RX I T X Z S RX RfH T X :dS RX g = m N RX N T X S T X 4p Z S RX Rf 1+g r r 3 e gr g(2 cosq) ˆ r+Rf 1+g r+g 2 r 2 r 3 e gr g sinq ˆ q :dS RX = m N RX N T X S T X 4p Z S RX Rf 1+g r r 3 e gr g(2 cosq)(ˆ r: ˆ n rx )dS RX + Rf 1+g r+g 2 r 2 r 3 e gr g sinq( ˆ q: ˆ n rx )dS RX (2.8) Observe that (2.1) and (2.2) have terms involving g and r. The terms with power 3 of 1=r are associated with the reactive near-field, and therefore the MI, and the terms proportional to 1=r are associated with the radiation (RAD) part of the signal. The inductive coupling is energy- efficient if the inductive part of the signal dominates the magnetic field [68]. Figure 2.3 provides the relationship between maximum range and operating frequencies for several MI/RAD ratios (q = 0) and shows that operating frequencies below 30 MHz can provide the required communication range of 2 meters for MI/RAD of 1. 15 (a) Old coordinate system (b) New coordinate system Figure 2.4: Relative location and alignment of the transmitter (TX) and receiver (RX) coils in different coordinate systems. Simplified Solution: assume that the TX coils is centered at origin with surface normal aligned in z direction, and RX coil is centered at C rx =(x 0 ;y 0 ;z 0 ) with a surface normal of ˆ n rx , as shown in Figure 2.4(a). The mutual inductance of two coils can be calculated by using (2.6) and the real parts of (2.1) and (2.2). Assuming an efficient inductive link (gr<< 1), we can write: M= m N RX N T X S T X 4p Z Rf 1+gr r 3 e gr g 2 cosq(ˆ r:ˆ n rx )+ sinq( ˆ q:ˆ n rx ) dS RX (2.9) Without loss of generality we can assume that the relative alignment of RX coil is in the form expressed as follows: ˆ n rx = ˆ yn y + ˆ zn z (2.10) So we have: M= m N RX N T X S T X 4p Z Rf 1+gr r 3 e gr g (3n z cos 2 q+ 3n y cosq sinq sinf n z )dS RX (2.11) Next, we define a new Cylindrical coordinate system (r 0 ;f 0 ;z 0 ) with an origin at the center of RX coil and ˆ z 0 = ˆ n rx , corresponding to (x 0 ;y 0 ;z 0 ), the Cartesian coordinate system shown in Fig- ure 2.4(b). Considering the coordinate transformation, the relationship between the old spherical 16 coordinate system (r;q;f) and the new Cylindrical coordinate system (r 0 ;f 0 ;z 0 ) is given by: r= p x 2 + y 2 + z 2 (2.12) q = Arccos( z r ) (2.13) f = Arccos( x p x 2 + y 2 ) (2.14) x= x 0 +r 0 cosf 0 (2.15) y= y 0 + n y z 0 + n z r 0 sinf 0 (2.16) z= z 0 + n z z 0 n y r 0 sinf 0 (2.17) In the new coordinate system, the integration over the surface of RX coil with an arbitrary location and alignment is transformed to an integration over a disc centered at the origin (z 0 = 0). To simplify the closed-form expression of mutual inductance, two functions f(r 0 ;f 0 ) and g(r 0 ;f 0 ) are defined such that: f(r 0 ;f 0 )=n z (x 2 0 + y 2 0 2z 2 0 )+ 3y 0 z 0 n y n z r 0 2 2x 0 n z r 0 cosf 0 (y 0 (n 2 y + 2)+ z 0 n y n z )r 0 sinf 0 (2.18) g(r 0 ;f 0 )=[ x 2 0 + y 2 0 + z 2 0 +r 0 2 + 2x 0 r 0 cosf 0 + 2(y 0 n z z 0 n y )r 0 sinf 0 ] 1=2 (2.19) Therefore, one can derive an expression for mutual inductance M as follows: M=m N RX N T X S T X Z 2p f 0 =0 Z a r r 0 =0 m(r 0 ;f 0 )r 0 dr 0 df 0 (2.20) and m(r 0 ;f 0 )= f(r 0 ;f 0 ) 4p Rf 1+g g(r 0 ;f 0 ) g(r 0 ;f 0 ) 5 e g g(r 0 ;f 0 ) g (2.21) where a r is the radius of RX coil, m(r 0 ;f 0 ) is the differential mutual inductance. 17 General Solution: any relative alignment and location of RX coil can be transformed into the form we used in the above expressions. Assume that the transmitter coil is centered at C T X , and its surface normal is ˆ n T X . The receiver coil is centered at C RX , and its surface normal is ˆ n RX . The more advanced version of the expressions without any simplification can be calculated by using the exact expressions of the magnetic field generated by the TX coil H T X as follows [34, 66]: M= m N T X N RX S T X 4p Z 2p f 0 =0 Z a RX r 0 =0 r df dr h Rf 1+gr+g 2 r 2 r 5 e gr g[r 2 cosar sinf (1+ cos 2 a)(c rx :ˆ y) r sinf sina cosa(c rx :ˆ z) 2r cosf cosa(c rx :ˆ x) cosa(c rx :ˆ x) 2 cosa(c rx :ˆ y) 2 sina(c rx :ˆ y)(c rx :ˆ z)] +Rf 1+gr r 5 e gr g[ cosa(c rx :ˆ z) 2 sina(c rx :ˆ z)(c rx :ˆ y) r sinf sin 2 a(c rx :ˆ y)+r sinf cosa sina(c rx :ˆ z)] i (2.22) where r is the distance between the origin and the observation point and can be defined in the cylindrical coordinates as follows: r= r(r;f)= h r 2 +(c rx :ˆ x) 2 +(c rx :ˆ y) 2 +(c rx :ˆ z) 2 + 2r cosf(c rx :ˆ x) + 2r sinf [cosa(c rx :ˆ y)+ sina(c rx :ˆ z)] i 1=2 (2.23) The parameters used in the above expressions are calculated from location and alignment of TX/RX coils as follows: a = tan 1 ( ˆ n rx :ˆ y ˆ n rx :ˆ z ) (2.24) ˆ n rx = R z (q z )R y (q y )R x (q x ) ˆ n RX (2.25) C rx = R z (q z )R y (q y )R x (q x )(C RX C T X ) (2.26) 18 q x = tan 1 ( ˆ n T X :ˆ y ˆ n T X :ˆ z ) (2.27) q y = tan 1 ( ˆ n T X :ˆ x p (ˆ n T X :ˆ y) 2 +(ˆ n T X :ˆ z) 2 ) (2.28) q z = tan 1 (ˆ n RX :ˆ x)[(ˆ n T X :ˆ y) 2 +(ˆ n T X :ˆ z) 2 ] (ˆ n RX :ˆ y)(ˆ n T X :ˆ z)(ˆ n T X :ˆ y)(ˆ n RX :ˆ z) (ˆ n T X :ˆ x)[(ˆ n T X :ˆ y) 2 (ˆ n RX :ˆ y)(ˆ n T X :ˆ z)(ˆ n RX :ˆ z)] (ˆ n RX :ˆ y)(ˆ n T X :ˆ z)(ˆ n T X :ˆ y)(ˆ n RX :ˆ z) (2.29) where R x (q x );R y (q y );R z (q z ) are rotation matrices that rotate vectors by an angle q x ;q y ;q z about the x-, y-, or z-axis using the right-hand rule, and are defined as follows: R x (q)= 2 6 6 6 6 4 1 0 0 0 cosq sinq 0 sinq cosq 3 7 7 7 7 5 (2.30) R y (q)= 2 6 6 6 6 4 cosq 0 sinq 0 1 0 sinq 0 cosq 3 7 7 7 7 5 (2.31) R z (q)= 2 6 6 6 6 4 cosq sinq 0 sinq cosq 0 0 0 1 3 7 7 7 7 5 (2.32) 2.2.2 Two-port Network Model Considering the mutual inductance between two coils and the equivalent circuit model of a loop antenna shown in Figure 2.5(a), the MI system can be modeled as Figure 2.5(b). The circuit model of a coil consists of self-inductance L, effective resistance R e f f , and parasitic capacitance C p . At low frequencies, the parasitic capacitance of a coil can usually be ignored, but it might be a major 19 (a) Loop antenna (b) Magnetic induction (MI) system with input and output matching networks Figure 2.5: Equivalent Two-port network model of MI system. problem in high-frequency circuits. As the operating frequency range used here does not exceed 50 MHz, the self-capacitance of MI coils is neglected. Therefore, a coil can be represented by its self-inductance and effective resistance, and consequently, the MI system can be modeled, as shown in Figure 2.5(b). The transmitter side of MI communication system has a source V s with impedance Z s at its input, and the receiver side has a load impedance Z L at its output. The parameters required for the circuit model are M;L T X ;R T X ;L RX ;, and R RX , which are the mutual inductance between coils, inductance, and resistance of transmitter and receiver coils, respectively. In order to facilitate performance analysis of the MI-based communication system around the human body, the closed- form expressions of these circuit parameters are presented [34]. The self-inductance of a coil with radius a, length b, number of turns N, circular cross-section wire, core-material permeabilitym, and wire diameterf w is calculated by the Coffin’s formula [69] as follows: L=m aN 2 [log 8a b 1 2 + b 2 32a 2 (log 8a b + 1 4 ) b 4 1024a 4 (log 8a b 2 3 ) + 10b 6 131072a 6 (log 8a b 109 120 ) 35b 8 4194304a 8 (log 8a b 431 420 )] (2.33) 20 The resistance of a coil comprises direct current (DC) resistivity, skin depthd w , and proximity effects. For a coil with wire resistivity ofr w it can be expressed as: R= 8 > > < > > : 2aNr w d w (f w d w ) iff w <d w 8aNr w f 2 w iff w d w (2.34) Suppose the current flowing through the TX coil is sinusoidal, then due to mutual inductance, a sinusoidal magnetic flux will be linked to the RX coil. A changing magnetic flux induces a voltage V ind in the RX coil [64]. Using the expression for M in (2.20), the induced voltage can be expressed as follows: V ind (t)=M d I T X dt = jw M I T X (2.35) 2.2.3 Lumped Reactive Impedance Matching To maximize the transmission efficiency of the overall system, each coil is attached to a frequency tuning network, called input and output matching networks [70]. Figure 2.5(b) shows the MI system with input and output matching networks. According to the maximum power transfer theorem, the input matching network must adjust the input impedance of the MI system Z in to the source impedance Z S , and the output matching network must match the load impedance Z L to the output impedance of the MI system Z out . The mutual inductance between the two coils is very small compared to the self-inductance of MI coils. Hence, we can approximate the input and output impedances by Z in = R T X + jwL T X (2.36) Z out = R RX + jwL RX (2.37) These circuits can be realized with lumped elements in three main types of L (normal and reversed), P, and T . as shown in Figure 2.6. 21 The simplest network is the L-network, which requires two reactive components. However, its Q factor and bandwidth are fixed by the values of the load Z L = R L + jX L , and the source impedance Z S = R S + jX S . The value of reactances used in normal and reversed L-matching network, X 1 and X 2 , can be obtained by Equations (2.38) and Equations (2.39), respectively [71]. X 1 = X S R S Q R S R L 1 X 2 =(X L R L Q) Q= s X 2 S + R 2 S R S R L 1 (2.38) X 1 = X L R L Q R L R S 1 X 2 =(X S R S Q) Q= s X 2 L + R 2 L R S R L 1 (2.39) TheP and T -networks require three reactive components, which provides the system with an extra degree of freedom to control the bandwidth. Here we discuss only theP matching network design, as it can be transformed to its equivalent T -network. TheP-network can be thought of as a normal network with two reactances X 1 ;X 4 , and reversed L-network consist of reactances X 3 ;X 5 , where X 2 = X 4 +X 5 . The normal L-section can be interpreted as a network that matches the source to the intermediate reference impedance Z= R+ jX. The reversed L-section can also be interpreted as a network that matches the intermediate reference impedance to the load [71]. In order to have a solution with the given Q-factor, which is determined by the application requirements, the resistance part of intermediate reference impedance must satisfy the condition 2.40. There are two solutions for each one of X 4 and X 5 that leads to four possible solutions for the reactances of the 22 Figure 2.6: Lumped reactive matching networks. P-network. R= max(R G ;R L ) 1+ Q 2 < min(R G ;R L ) (2.40) 2.2.4 Power Gain The power gain of a system is defined as the ratio of received power by the load to the transmitted power and is proportional to the square of the forward voltage gain magnitude S 21 . G=jS 21 j 2 (2.41) The S-parameters, which are among the most commonly used parameters for performance analysis, show the relationship between the reflected and incident power waves. As the proposed MI communication system is a cascaded connection of two-port networks, the ABCD parameters are the best candidate for performance analysis of the whole system and calculate its power gain. 23 These parameters, which are also known as transmission, chain or cascade parameters, relate the input current and voltage at port 1 to the output. The ABCD parameters of the MI system are equivalent to the product of ABCD matrices corresponding to the input matching network, the MI system, and the output matching network, accordingly. 2 6 4 A B C D 3 7 5 = 2 6 4 A im B im C im D im 3 7 5 : 2 6 4 A mi B mi C mi D mi 3 7 5 : 2 6 4 A om B om C om D om 3 7 5 (2.42) where A;B;C;D are the ABCD parameters of the overall MI system, including the MI coils and the matching circuits, and A ;B ;C ;D are the ABCD parameters of input matching network (im), the MI system (mi), or the output matching network (om). The S-parameters of the proposed MI circuit model can be calculated by using ABCD param- eters to S-parameter conversion. Applying the ABCD parameters to S-parameters conversion, the forward voltage gain of the MI system S 21 can be determined as follows [72]: S 21 = 2 p RfZ S gRfZ L g AZ L + B+C Z S Z L + DZ S (2.43) 2.3 Model Validation Verification and validation of the proposed model via simulation and experiment are discussed in this section. The finite-element method (FEM) based Ansys high frequency structural simulator (HFSS) software is used as a numerical simulator, shown in Figure 2.7. For the experimental mea- surements, an Agilent vector network analyzer (VNA) is used, and a picture of its setup can be seen in Figure 2.8. To verify the assumption that the effect of the human body on MI signal is neg- ligible, experiments are performed with and without the presence of the human body. Moreover, several simulations and measurements are performed to study the effect of the impedance match- 24 Figure 2.7: HFSS simulation setup for model validation. Figure 2.8: Measurement setup for model valida- tion. ing networks on the MI system. As one of the key performance metrics for evaluating a wireless communication system is the power gain, here we measure the S-parameters of the MI system with and without impedance matching. 2.3.1 Experimental Setup The MI system used in simulations and measurements consists of two identical air-cored, single layer copper coils in the background medium of air. The experiments are performed for various coils, summarized in Table 2.1, in different locations and alignments relative to each other. We assume that the TX coil is centered at the origin, and its surface is in the xy-plane. The RX coil is centered at the defined location (C RX ), and its surface is parallel to the q x;RX degree rotated version of xy-plane about the x-axis. The relative location of these coils is defined by their distance d, lateral misalignmentD, and angular misalignmentq. The S-parameters of the MI system are measured in the frequency range of 1 MHz to 50 MHz. Here, both the input and output impedances are the same and are considered to be 50W (Z 0 = 50). Therefore, the load impedance Z L , and source impedance Z S in Figure 2.5(b) are considered to be the same as the characteristic impedances (Z L = Z S = Z 0 ). 25 Figure 2.9: Magnetic induc- tion (MI) node prototype. Figure 2.10: Measurement setup with matching circuit for model validation. We have also verified the simulation results of the MI system with different matching networks via experimental measurements. The MI node, including a coil and an L-matching circuit used for experiments, is shown in Figure 2.9. The VNA setup is shown in Figure 2.10. The coils with 5 cm radius and 10 American wire gauge (AWG) wire diameter are considered as the TX/RX antennas (case #3), and the resonance frequency is 13.56 MHz. Resistance and self-inductance of the MI coils measured by VNA at the resonance frequency are 101 mW and 241 nH, respectively. The matching network is designed by the equations and conditions reviewed in section 2.2.3. Two matching networks are designed to tune the impedance of MI coils to Z 0 : reversed L- matching, andP-matching circuit. The reversed L-matching network consists of a series inductor L 2 = 5380 nH, and a parallel capacitor C 1 = 600 pF. The P-matching network consists of two parallel inductors L 1 = L 3 = 19:6 nH, and a series capacitance, C 2 = 3:5 nF. All the parameters of the tuning circuits satisfy the matching conditions. We have also compared the results with the Table 2.1: The magnetic induction (MI) coils parameters used for model validation. case number #1 #2 #3 #4 #5 #6 #7 #8 coil radius - cm (a) 2 4 5 5 5 5 7 3 number of turns (N) 2 1 1 1 1 1 1 2 wire gauge (AWG) 18 18 10 12 16 18 12 10 26 Figure 2.11: Dielectric properties of human body tissues. resonant MI system. The resonant MI system includes a series capacitance C s of the coil such that(2p f res ) 2 LC s = 1 ( where L, and f res denote the self-inductance of coil, and the resonance fre- quency of MI system, respectively) [34]. Hence. the value of series capacitance used for simulation of resonant MI system is C s = 574 pF. 2.3.2 Human Body Model and Dielectric Properties The Ansoft model used for the human body is an adult male with more than 300 objects represent- ing internal organs. Most biological tissues have permeability close to air [73], and their dielectric properties are determined from the complex relative permittivitye . The relative permittivity and electrical conductivity of several tissue types used for the human body are shown in Figure 2.11, based on the results presented in [74]. Complex relative permittivity is modeled with the summation of 4-Cole-Cole dispersion re- gions [74]. The dispersion regions characterize the behavior of tissues at low, medium, and high frequencies, which are referred to asa,b, andg dispersion, respectively. The expression describes the variation of dielectric properties as a function of frequency in the range of 10 Hz to 100 GHz. 27 e (w)=e 0 je 00 =e ¥ + å n 4e n 1+( jwt n ) (1a n ) + s i jwe 0 (2.44) where t n is the relaxation time, s i is the static ionic conductivity, a n is the distribution parameter which is a measure of the broadening of the dispersion (between 0 and 1). The parameter4e n is the magnitude of dispersion which is calculated as follows: 4e n =e s e ¥ (2.45) wheree s is the permittivity at static wherewt<< 1, ande ¥ is the permittivity at infinite frequency wherewt >> 1 [75]. 2.3.3 Results Figure 2.12 displays the results of experiments and simulations for the scenario where both MI coils are perfectly aligned and separated by a distance d. The vertical axis of the plots shows the magnitude of the forward voltage gain in decibels, and the horizontal axis shows the operating fre- quency range. In the second scenario, MI coils are fixed in their defined locations but have differ- ent alignment relative to each other. Figure 2.13 presents results of simulations and measurements where two coils separated by 40 cm distance, and the RX coil’s surface rotates around the x-axis. It means that the location of the RX coil is fixed, and angular misalignment (q) varies between 0 to 180 degrees. The simulation and experimental measurements for MI coils with matching networks in different distances, lateral misalignment, and angular misalignment are also presented in Figure 2.14. 28 (a) a = 5cm, N = 1, AWG = 10 (b) a = 7cm, N = 1, AWG = 12 Figure 2.12: The power gain of two perfectly aligned circular coils separated by relative distance of d in the background medium air. 29 (a) a = 5cm, N = 1, AWG = 10 (b) a = 7cm, N = 1, AWG = 12 Figure 2.13: The power gain of two circular coils separated by distance of 40 cm, and surface misalignment of q in the background medium air. The location of transmitter (TX) and receiver (RX) coils are fixed and angular misalignment (q) varies between 0 to 180 degrees. 30 (a) D = 20cm,D = 0cm,q = 0 (b) D = 30cm,D = 10cm,q = 0 (c) D = 30cm,D = 0cm,q = 45 Figure 2.14: Simulation and experimental measurement results. The power gain comparison of magnetic induction (MI) system with and without matching networks. 31 2.3.4 Discussion The results of the MI system without matching networks show that the accuracy decreases with increasing distance between the RX and TX coils. The reason is that the MI/RAD ratio is in- versely proportional to the square of the distance. As the RX coil gets farther, the inductive part extinguishes very rapidly, and the RAD part dominates the MI part of the signal. The propagation medium can affect the radiation part of the signal and result in an increased model error. We can conclude that the proposed model is accurate in the near-field and mid-field regions. The boundary of these regions depends on the dominant wavelength emitted by the source and the size of the RX coil. The results of MI system with matching networks show that the resonance frequency is at the close proximity of the desired frequency, and impedance matching has improved the system efficiency significantly compared to the resonant technique. The difference between simulation and measurement results are due to the strong dependency of resonance frequency on the exact value of L T X ;L RX ;C s . Hence, any small deviation from the designed values can change the reso- nance frequency and cause a significant gain drop. This deviation is more pronounced for larger Q-factors. Moreover, any change in temperature, capacitor type, and coil geometry can also lead to a deviation. In order to evaluate the proposed model, the error for each type of coil shown in Table 2.1 is calculated for several different locations and alignments, and their normalized root-mean-squared error (NRMSE) is expressed as a function of frequency and angular misalignment. The NRMSE of the estimated values ( ˆ jS 21 j) and the observed values (jS 21 j) of MI system without matching networks for different predictions is defined as: NRMSE= RMSE E[jS 21 j] (2.46) where root-mean-squared error (RMSE) is defined as follows: RMSE= s å N i=1 (jS 21 jj b S 21 j) 2 N (2.47) 32 (a) HFSS results without the human body. (b) HFSS results with the human body. (c) Experimental measurement results without the human body. (d) Experimental measurement results with the human body. Figure 2.15: The NRMSE of estimated forward voltage gain with respect to simulation and exper- imental measurements vs. frequency (left), andq x (right). 33 where the operatorE[:] indicates the expected value over all possible samples. As Figure 2.15 shows, the difference between the measured values and predicted by the model is very small, and their NRMSE for frequencies between 1 MHz to 30 MHz is below 15%. For higher frequencies, any small parasitic capacitance, such as probe capacitance, forms a resonant circuit and results in higher errors. The NRMSE of the model with respect to the simulation results without the presence of the human body is below 10%. For frequencies above 30 MHz, the NRMSE of the model with respect to the HFSS simulations in the presence of the human body quickly rises above 20%, which is no longer negligible. This is due to the fact that the biological tissues are non-magnetic, and the human body does not affect the magnetic field for frequencies below 30 MHz. Since the magnetic field is dominant in the near field of the loop antenna, this antenna is robust in performance near the human body. Moreover, the wavelength of frequencies below 30 MHz is considerably larger with respect to the dimension of the human body, and as a result, the human body and the associated losses do not affect the performance of MI system. The gain computation in the presence of the human body by the MI model is significantly faster than computation using HFSS, allowing rapid calculations in designing MI-WBANs, and assessing their performance over a large range of parameters. The latter point was a major motivation behind the work reported here. 2.4 Model Accuracy Analysis in Generating Time-series Data We have studied the effect of coils’ spatial translation and rotation on the MI signal using exper- imental measurements. The MI signals measured during coils movement are then compared with the synthetic MI signals corresponding to the captured motion for evaluating the model accuracy in generating time-series MI data. For these experiments, we used two identical, air-cored, single layer copper coils with 5 cm radius, 10 AWG wire diameter. The forward voltage gain of two in-motion coils is measured for 30 seconds via a VNA with 1800 points resolution. The corresponding synthetic S 21 is also generated using the system model for comparison. All parameters of the model are predefined based on the MI system setup except 34 Figure 2.16: Schematic representation of measurement setup. the distance and misalignment between coils, which are variable during the movement. Hence, two coils are labeled with red markers and placed in front of a green screen. The motion of coils is captured via an iPhone’s built-in camera with 30 fps, and the videos are processed offline to extract markers, their center, and alignment. Figure 2.16 and 2.17 depict the measurement setup and a camera frame sample after video processing, respectively. The object tracking is applied to extract red markers attached to coils and the blue fixed-length calibration label. The coils only move in 2D without losing generality, such that the camera can capture their motion as only one camera is used. The extracted pixel-wise movement of coils is then converted to the spatial translation using a predefined length ‘calibration label’. The ratio of the calibration label’s length to its size extracted from video provides a meter to pixel ratio. As the camera is fixed during the experiment, this ratio remains constant for all frames of the video. The recorded distance between coils covers up to 60 cm range. The recorded video and VNA measurements are not synchronized. Hence, the generated synthetic MI data needs to be synchronized with measured data by minimizing the NRMSE. Figure 2.18 shows the measured and simulated (synchronous and asynchronous) forward voltage gain of two coils during different movement. We have performed experiments for 20 different motions that involve both geo-translation and misalignment of coils. The Results show that the simulated signal is consistent with the measured data, which is an indication of a valid model for generating time-series MI-data. The average 35 NRMSE of the synthesized and measured S 21 for these experiments is less than 10.3%. The re- ported NRMSE not only takes into account model error but also includes the error associated with the motion tracking algorithm using video and VNA measurements. Figure 2.17: A camera frame sample after processing. 2.5 Conclusion We studied the performance of MI communication systems with and without the presence of the human body. The mutual inductance between coils is a key parameter in computing the gain of MI communication system and power transmission analysis. Using the derived mutual inductance expression, the MI system is modeled as a circuit. The received voltage at the secondary coil, and as a result, the forward gain voltage can be calculated. Model validation is performed using simulations and experiments. The computed results from the model compare well with the results of HFSS simulations and measurements. Gain calculation by the proposed model is significantly faster than computation using the full-wave simulator. The HFSS simulator is also used to compute the forward gain voltage between two coils with the presence of the human body. Results have shown that the MI model is valid for the operating frequency of up to 30 MHz, where the dimension of the human body is relatively small compared to the wavelength, and the effect of the human body on the performance of MI system is negligible. Higher frequencies are beyond the magnetic induction regime, for which propagation and scatter- 36 Figure 2.18: Measured vs. synthetic magnetic induction (MI) data. The measured and simulated voltage gain of two MI coils during arbitrary movements, such that both relative alignment and location of coils vary. 37 ing effects become significant. In that range, the human body affects the signal, and the model is not valid anymore. The proposed closed-form expression for mutual inductance considers the effect of frequency, coil geometry, misalignment, distance, and environment. These parameters should be addressed for MI communication system design, which means that a design strategy for MI-based WBAN is also required. 38 Chapter 3 Human Activity Recognition 3.1 Introduction The objective of human activity recognition (HAR) is to provide information on human physical activity and detect simple or complex actions in a real-world setting. It allows computer systems to assist users with their tasks and to improve the quality of life in areas such as senior care, re- habilitation, daily life-logging, personal fitness, and assistance for people with cognitive disorders [4, 30, 76–79]. Two main approaches for deployment of HAR systems are external, and wearable sensors [29]. In the external approach, the monitoring devices are set at fixed points, and users are expected to interact with them [80]. For example, the Vision-based technique is one of the well- known external methods that has been extensively studied for human activity analysis [81, 82]. However, it faces many challenges in terms of coverage, accuracy, privacy, and cost. It requires in- frastructure support, such as installing video cameras in surveillance areas, which is usually costly. Additionally, cameras cannot capture any data if the user performs out of their reach [83, 84]. In the second approach, on-body sensors, such as accelerometers, gyroscopes, and magnetome- ters, are used to translate human motion into signal patterns for activity recognition [54, 85, 86]. Recent advances in embedded sensor technology have made it feasible to monitor the user’s ac- tivity using smart devices. Several research studies have reported the use of smartwatches and smartphones in human activity monitoring and have presented a satisfactory performance [13–16]. 39 Although these devices provide a privacy-aware alternative solution that overcomes many disad- vantages of the external approach, they still might not be able to address the requirements of a diverse range of applications. A single wearable cannot cover the entire body and therefore fails to obtain adequate informa- tion about the mobility of all body segments [19, 20, 22]. For example, inertial sensors embedded in a smartwatch cannot capture the movement of legs, which restricts the capability of the system in classifying activities. Additionally, in systems relying on data from a single device, variations in position can significantly affect the performance or lead to the failure of the monitoring system [19, 21, 23]. WBAN consisting of wearable devices operating around the human body can tackle these problems [22, 24]. In WBANs, sensors are spatially distributed over the human body and collect data from the user. Then data are transmitted wirelessly to a central processing unit for de- tection. This approach can provide comprehensive information on the mobility of body segments and potentially improve system accuracy. WBAN design is, however, challenging as many constraining, often conflicting, requirements have to be taken into account [25–27]. For example, the system must be inexpensive, accessible to the general public, and meet ergonomic constraints and health requirements. It has to operate under proper guidelines limiting the power exposure to the user since the energy absorption may lead to temperature elevation in biological tissues. To ensure users’ safety, it has to satisfy SAR constraints while providing a reliable wireless link [28]. Moreover, the system should guarantee the security and privacy of the user’s data. Wearable devices must be small and lightweight, which puts a restriction on the battery size and longevity. On the other hand, frequent battery recharging may not be practical for sensor networks with multiple sensors in applications such as senior monitoring [29]. Due to energy resources limitation, the power management has become a critical issue in de- signing a WBAN. Since wireless communication consumes a considerable portion of the energy [31], numerous studies have proposed and investigated low-power solutions [32–35]. The conven- tional state-of-the-art wireless sensor networks working in the vicinity of the human body adopt 40 Figure 3.1: Magnetic induction-based human activity recognition (MI-HAR) system. radio wave propagation for signal transmission. This technique is susceptible to the characteristics of the environment, and its signal experiences a high attenuation around a lossy medium, such as the human body. It results in higher power consumption, shorter battery life, and lower reliability [34, 38, 39]. Moreover, radio-wave propagation technologies are prone to interference with adja- cent communication links since most of them, such as Bluetooth, operate at the busy 2.4 GHz, the IMS band [40, 41]. They also have potential security problems as their signal cannot be stopped from propagating into free-space. Therefore it can be intercepted even distant from the transmitter [43]. We introduce the MI-HAR system that effectively detects physical movements by MI signals (Figure 3.1). This system represents the motion of human body parts via variations in the MI signals transmitted from transmitter to the receiver during physical action, instead of spatial data measured by the inertial sensors. This approach can overcome several problems associated with conventional sensor-based HAR systems, such as eliminating the need for an extra wireless mod- ule, reducing power consumption, and the required bandwidth by combining data collection and wireless signal transmission steps. The manufacturing cost of an MI module is approximately less than $20, while a Bluetooth inertial measurement unit (IMU) costs more than $100 [87–89]. Moreover, it has other features that are inherited from the MI-based communication system. 41 This chapter is organized as follows: Section 3.2 explains the system principle, and Section 3.3 presents the framework, including two stages of data acquisition and detection. The process of synthesizing MI motion data corresponding to several physical activities is described in Section 3.4. The machine learning-based classifiers and deep recurrent neural networks for the classifica- tion of human movements and their results are presented in Section 3.5, and Section 3.6 concludes the results. 3.2 System Principle The MI-based communication system is a short-range wireless physical layer that transmits sig- nals by coupling a non-propagating magnetic field between the wire coils rather than radiating as conventional methods. As explained in chapter 2, the signal generated by an MI coil attenuates as a function of frequency, channel medium, coils’ geometry, location, and alignment [34]. The non-propagating magnetic field is mainly affected by the permeability of the medium, which is close to the air for non-ferrous materials. The MI channel condition remains constant even in an inhomogeneous lossy medium, such as around the human body [34, 62]. For the frequency of up to 30 MHz, the dimension of the human body is relatively small com- pared to the wavelength, which makes the propagation and scattering effects insignificant [34]. The immunity of signal in this frequency range to the environment makes the forward voltage gain, S 21 , of the MI system only a function of coils’ locations and alignments for a predefined coil geometry and operating frequency. The gain varies by changing the distance and alignment between the MI coils, and therefore, relative motion between the MI coils yields patterns in the received MI signal. This unique characteristic of the MI system is the fundamental principle of the proposed MI-HAR system. 42 3.3 Framework The activity recognition process steps are different depending on the application. The framework used here has two main stages: data acquisition and detection. For the first stage, an MI-based communication system is employed, which enables the integration of sensing and wireless data transfer into a single step. The user wears the RX coil, for example, as a belt around the waist, and TX coils can be placed around the other skeleton bones, such as wrists, arms, and legs. The human body bones are spatially translated and oriented during a physical activity, which changes the relative location and alignment of the MI coils around them. Collecting the received MI signals transmitted from the coils enclosing skeleton bones can model the relative motion of human bones to represent motion. Since the spatial variations of skeleton bones over time are discriminative descriptors of human actions [90], the vector of sam- ples observed by the MI coils over time can be considered as the set of inputs for the activity detection algorithm. Increasing the number of coils around the skeleton bones results in a broader set of input data. It consequently enhances the accuracy of the MI-HAR system in detecting the relative motion of body parts. In the next step, a classification method is applied to the MI motion data for detecting human action. 3.4 Synthetic Data We synthesized MI motion data to evaluate the proposed MI-HAR system capability in motion detection. The circuit model of the MI system presented in section 2.2.4 is used to calculate the forward voltage gain, which is the scaled version of the received MI signal. As the pattern is the same, we used the generated voltage gain patterns of the system as the input features for the detection algorithm. To synthesize MI motion data during different human actions. We considered a receiver and eight transmitter coils around the torso, hands, arms, legs, and thighs, respectively. The setup is explained in detail in section 3.4.1. For spatial translation and rotation of human body 43 Figure 3.2: Location of MI transceivers on the human body. Figure 3.3: Location of markers. bones, 3D motion capture (MoCap) datasets are employed. Each pair of markers placed at the joints can define a bone. Hence, the location and alignment of MI coils placed around the body parts can be derived and provided as inputs to the model for synthesizing the corresponding MI motion data. 3.4.1 Magnetic Induction System Setup The MI transceivers adopted in the experiments consist of a coil and L-reversed impedance match- ing network [70]. The matching network is used to maximize the transmission efficiency of the overall system [70]. The coils are identical, air-cored, single layer copper with 5 cm radius, 10 AWG wire diameter, and the user can wear them as accessories. The coil’s radius can change depending on the size of the body part that they are designed to be placed around. The source and load impedances are 50W, and the resonance frequency is 13.56 MHz. As the operating frequency is lower than 30 MHz, the human body effect is neglected [34], and the effect of the background medium is considered to be the same as that of air. The reversed L-matching networks consist of a series inductor of 5380 nH and a parallel capacitor of 600 pF. 44 Figure 3.2 depicts the location of MI transceivers (TX i , RX) considered around the human body and the laboratory version of an MI transceiver for generating synthetic MI motion data. The location of markers in 3D (M j ) on the human body required to track coils motion is also displayed in Figure 3.3. The human body parts, including the torso, left arm, left hand, left thigh, left leg, right arm, right hand, right thigh, right leg, and their corresponding marker pairs are summarized in the table 3.1. Table 3.1: The marker pairs defining two ends of a body part. Human Body Part Markers Torso (M 2 ; M 9 ) Left Arm (M 3 ; M 4 ) Left Hand (M 4 ; M 5 ) Left Thigh (M 10 ; M 11 ) Left Leg (M 11 ; M 12 ) Right Arm (M 6 ; M 7 ) Right Hand (M 7 ; M 8 ) Right Thigh (M 13 ; M 14 ) Right Leg (M 14 ; M 15 ) Consequently, these pairs can be utilized to calculate the location of coils (C T X i , C RX ) and their alignment ( ˆ n T X i , ˆ n RX ). Assuming that the coils are located at the midpoint of bones, we can calculate their center by averaging the location of the corresponding pair of markers. For example, Figure 3.4 shows the right leg, its corresponding transmitter coil, and markers. It depicts how the center and alignment of a bone and its corresponding coil T X 8 can be calculated using markers locations as: c T X 8 = M 14 + M 15 2 (3.1) 45 Figure 3.4: The center and alignment of a bone and its corresponding coil that can be calculated using markers locations. The coils are around the human bones, which indicates that the alignment of the line passing through the markers is the same as the surface normal of its corresponding coil. Therefore, the surface normal of the transmitter T X 8 can be written as: ˆ n T X 8 = M 15 M 14 jM 15 M 14 j (3.2) 3.4.2 Motion Capture Datasets Two publicly available experimental datasets are used in this dissertation, containing diverse move- ment data to verify the applicability of MI-HAR in detecting a wide range of activities. A brief description of these datasets is as follows: 3.4.2.1 Biological Motion Library (BML) Biological Motion Library (BML) dataset [91] contains a full-body movement dataset for four activities including walking, knocking, lifting, and throwing performed by 15 male and 15 female actors in a neutral, angry, happy, and sad style. The dataset is balanced and has the same number of 46 records performed by actors for each action. The total number of samples is 1028, with a sampling rate of 60 Hz. For walking action, the data are captured for 30 seconds of walking in a triangle turning rightward (clockwise) and turning leftward (counterclockwise). For the knocking, lifting, and throwing actions, five repetitions of a single action unit are obtained for each data record, approximately 20 seconds in duration. 3.4.2.2 Multimodal Human Action Database (MHAD) Multimodal Human Action Database (MHAD) dataset [92] contains the data for 11 activities in- cluding jumping, jumping jacks, bending, punching, waving two hands, waving one hand, clap- ping, throwing, sit down/stand up, sit down, stand up. The actions are performed by 7 male and 5 female subjects in the range of 23-30 years of age except for one elderly subject. The number of records for each action is the same, yielding about 660 data sequences correspond to a total record- ing time of 82 minutes. Except for sitting down, stand up, and throwing, all records include five repetitions of a single action. The approximate recording length of activities varies from 2 to 15 seconds. The dataset consists of data from four microphones with a sampling rate of 48 kHz; six accelerometers fixed on wrists, hips, and ankles with a sampling rate of 30 Hz, the optical motion capture system with a sampling rate of 480 Hz, cameras with a sampling rate of 22 Hz, and depth sensors with a sampling rate of 30 Hz. In our experiments, we used the downsampled MoCap data to 60 Hz. 3.4.3 Results The generated synthetic forward voltage gain of the MI transceivers corresponds to the BML and MHAD datasets are presented in Figure 3.5, and Figures 3.6 and 3.7 respectively. A point to con- sider is that we have extended the single-transmitter/single receiver model to a multi-transmitter/single- receiver scenario, assuming the interferences such as cross-coupling between coils are negligible because the interference mitigation techniques such as time-division multiplexing (TDM) [93] or frequency splitting [94] can be applied to reduce or ideally eliminate interference between in- 47 Figure 3.5: Synthetic magnetic induction (MI) motion data. The forward voltage gain S 21 between the receiver (RX) and transmitters (TX 1 -TX 8 ) are generated using the proposed MI model and the human motion data captured for different activities in Biological Motion Library (BML). 48 Figure 3.6: Synthetic magnetic induction (MI) motion data. The forward voltage gain S 21 between the receiver (RX) and transmitters (TX 1 -TX 8 ) are generated using the proposed MI model and the human motion data captured for different activities in Berkeley Multimodal Human Action Database (MHAD). 49 Figure 3.7: Synthetic magnetic induction (MI) motion data. The forward voltage gain S 21 between the receiver (RX) and transmitters (TX 1 -TX 8 ) are generated using the proposed MI model and the human motion data captured for different activities in Berkeley Multimodal Human Action Database (MHAD). 50 ductive systems. Moreover, interference protocols (e.g., radio frequency identification (RFID) interference protocols) can control communication between transceivers while preventing their in- terference with one other. Therefore, the model can provide a reasonably accurate estimation of multi-coil system performance. 3.5 Data Analysis 3.5.1 Data Preprocessing In our experiments, we have used the magnitude of MI signals as input for the classifiers. Data samples are processed before fetching into the classification models. For data cleaning, the miss- ing values are substituted with previous non-missing values, and a 5-point quadratic (order 4) polynomial Savitzky-Golay filter is applied for denoising. Then the baseline offset is removed from time-series data. In the MHAD dataset, 3% of the signals are removed from the end of each data sample as the reported experiments show improvement in the accuracy [95]. 3.5.2 Classification To assess the performance of the proposed MI-HAR system in recognizing human activities, we implemented deep recurrent neural network (RNN)s based on long short-term memory (LSTM) units due to their strong performance in human activity detection and their capability in learning complex representations of the motion data [96, 97]. We compared the results of this method with several commonly used classifiers for activity detection using generated synthetic MI motion data, presented in section 3.5.3. The models are trained and evaluated on the generated synthetic motion datasets of eight bones using the leave-one-subject-out cross-validation (LOSO-CV) method. For the experiments on the BML and MHAD dataset, six and two subjects are used for validation and the rest for training, respectively. 51 Figure 3.8: The process of calculating the bag-of-words representation of time-series motion data samples. 3.5.2.1 Bag of Words (BoW) We used the bag-of-words (BoW) representation to characterize the time-series data with different lengths. The flowchart of this process is presented in Figure 3.8. First, the synthetic MI motion data are divided into fixed-length segments of 1 second using the sliding window technique with 0.8 second overlap. Attributes are then computed for the time domain, frequency domain, and time-frequency domain of each window segment. Frequency domain and time-frequency domain representations of the signal are calculated by the fast Fourier transform (FFT) and single-level discrete Wavelet transform (DWT) based on the Daubechies2 wavelet filter, respectively. The attributes considered here are extremes, mean, median, standard deviation, lower quartile, upper quartile, skewness, kurtosis, and the correlation between each pair of signals. As each action is associated with eight data samples, the resulting feature vector for each segment is generated by the combination of eight feature sets. Features are also scaled using the min-max scaling method to bound values in the range of 0-1. The scaling makes the weight of all features equal in the process of classification. Next, the feature vectors from the training data are clustered using k-means clustering to define a codebook that contains the cluster centers, which are called codewords. Then, each window segment is assigned the closest codeword, and a time-series 52 is represented as a histogram of codewords. The BoW representations of synthetic MI motion data are used as inputs for the machine learning-based classification models. In our experiments, we quantized the training data of BML and MHAD datasets to 100 and 20 codewords, respectively. K-mean clustering Clustering is the task of identifying groups, called clusters, and structures in the data so that sam- ples of a group are in some sense more similar to each other than to those in other groups. It has applications in many fields, such as image analysis, pattern recognition, bioinformatics, data compression, etc. K-Means is one of the most popular clustering algorithms that aims to partition data points into k clusters. It identifies k centroids and uses them to define clusters. Then assigns each data point to the cluster with the nearest centroid. K-means iteratively loops over these two steps to cluster data. The k-means clustering algorithm is as follows [98]: 1. Randomly initialize K cluster centers m j . 2. Calculate the distance between each data point x i and cluster centers. Then assign it to the nearest cluster center according to the calculated distances. c i = argmin j jx i m j j 2 (3.3) 3. Recalculate the new cluster center using: m j = å N i=1 I(c i = j)x i å N i=1 I(c i = j) (3.4) whereI(:) is an indicator function, and N is the number of data points. 4. Repeat steps 2 and 3 until convergence (no changes in assignments). The algorithm is not guaranteed to find the optimum. 53 3.5.2.2 Machine Learning-based Classification Classification is the process of categorizing and predicting the label of given data points. It uses the training dataset to approximate a mapping function from input variables x to discrete output vari- able y, i.e., y2f1;:::;Cg with the number of classes C. Therefore, the training dataset must have sufficient samples of each class to represent the problem. For C= 2, it is called binary classifica- tion, and for C> 2, this is called multi-class classification. When the class labels are not mutually exclusive, and each sample is assigned to a set of labels, it is called a multi-label classification [99]. Several most common real-world examples of classification problems are medical diagnosis, speech recognition, document classification, face detection, handwriting recognition, etc. Different performance metrics are available to evaluate classification algorithms. Here we use accuracy, precision, recall, and F1-score. Accuracy is equal to the number of correct predictions to the total number of cases examined, which can be defined as follows: Accuracy= T P+ T N T P+ FP+ T N+ FN (3.5) where T P;T N;FP; and FN are True Positive, True Negatives, False Positives, and False Negatives, respectively. Precision shows the ratio of the instances classified in the positive class that is correctly classi- fied and is defined as follows: Precision= T P T P+ FP (3.6) Recall is defined as the ratio between all the correctly classified instances in the positive class against the total number of actual members of the positive class. IT shows the total numbers of positive samples that are correctly classified and can be written as follows: Recall= T P T P+ FN (3.7) 54 Increasing precision reduces false positives, and increasing recall reduces false negatives. Usu- ally, increasing the precision results in a decrease in recall and vice-versa. F-Measure or F-Score combines these two metrics into a single measure to consider both aspects. The F1 score is the har- monic mean of Precision and Recall provides a measure of the incorrectly classified cases defined as follows: F1= 2 Precision:Recall Precision+ Recall (3.8) Here we implemented The several popular machine learning-based classifiers using python library Sklearn [100]. The multi-class models are non-linear SVM with a polynomial kernel, K- nearest neighbors (KNN), decision trees (DT), random forest (RF), and logistic regression (LR). A brief introduction to these models are presented here. K-Nearest Neighbors KNN is a non-parametric memory-based learning method that stores multidimensional training data samplesD along with their labels y. It allows the classifier to classify new input x by assigning the most frequent label among the K training samples nearest to the given point. The parameter K is a user-defined constant. It can be formulated as follows [99]: p(y= cjx;D;K)= 1 K å i2N k (x;D) I(y i = c) (3.9) where N k (x;D) indicates the K nearest point to x inD, andI(:) is the indicator function defined as: I(e)= 8 > > < > > : 1 if e is true 0 otherwise (3.10) Different distance metrics can be employed to find the nearest neighbors, and the optimal one should be selected according to the dataset and classification problem. Hamming distance is a metric used for the categorical variables. It measures the number of instances at which the symbols are different among two strings of equal length. Euclidean distance, Manhattan, and Minkowski 55 are commonly used metric for continuous variables calculating the distance between two points X and Y in N-dimensional space as follows: Euclidean(X;Y)= s N å i=1 (x i y i ) 2 (3.11) Manhattan(X;Y)= N å i=1 jx i y i j (3.12) Minkowski(X;Y)= N å i=1 jx i y i j p ! 1 p for p> 1 (3.13) Another well-known metric is the Cosine similarity measuring the difference in direction between two vectors A and B instead of calculating the magnitude. similarity= cosq = A:B jAjjBj (3.14) Logistic Regression Logistic regression is a simple statistical classification algorithm that predicts the probability of a categorical target. The output variable is generally binary in nature (binomial logistic regression), but variables with more than two categories can also be predicted (multinomial or ordinal logistic regression). The logistic regression uses the linear regression to predict a value, which can be any real number. However, the output value should be squashed into a range of [0,1] to be interpreted as a probability [99]. Therefore, it is passed through the Sigmoid function, which is also known as logistic or logit function, defined as: sigm(h), 1 1+ e h = e h 1+ e h (3.15) 56 The logistic regression mathematical model for binary classification, where y2f0;1g, can be derived as: p(y= 1jx;w)= 1 1+ e w T x (3.16) Support Vector Machines The SVM algorithm aims to find a maximum marginal hyperplane (MMH) in N-dimensional space that distinctly classifies the data points. The closest points to the decision hyperplane are called support vectors, and the SVM objective is to maximize their distance to the hyperplane. It en- ables the model to classify new data points with more confidence. For the non-linearly separable dataset, the SVM uses kernel technique (e.g., linear, polynomial, and radial basis function (RBF)) to transform the input into a higher-dimensional space. Considering the two-class classification problem using linear models of the form y(x)= w T (x)+ b (3.17) where(x) denotes a fixed feature-space transformation, and b is the bias parameter. The training data set comprises N input vectors x 1 ;:::;x N , with corresponding target values t 1 ;:::;t N where t n 2 f1;+1g, and new data points x are classified according to the sign of y(x). Assuming that the training data set is linearly separable in feature space so at least one solution exists for w and b such that [98]: y(x n )> 0 for data points with t n =+1 (3.18) y(x n )< 0 for data points with t n =1 (3.19) so that t n y(x n )> 0 for all data points (3.20) 57 The perpendicular distance d of a point x from a hyperplane defined by y(x)= 0, where y(x) is presented in (3.18), can be calculated as follows: d= jy(x)j w (3.21) Considering all data points are correctly classified so that t n y(x n )> 0 for all n. Hence, the distance of a point x n to the decision surface is given by: t n y(x n ) jwj = t n (w T (x)+ b)) jwj (3.22) To maximize the given distance, the parameters w and b should be optimized. Thus the maxi- mum margin solution can be found by solving: argmax w;b 1 jwj min n [t n (w T (x)+ b)]] (3.23) For the point closest to the surface, we can consider t n (w T (x)+ b)= 1. Therefore, all data points satisfy the constraints: t n (w T (x)+ b)> 1 n= 1;::;N (3.24) The data points that satisfy equality constraints are called active, while others are inactive. By definition, there are always at least two active constraints when the margin is maximized. The optimization problem then tries to maximize 1=jwj, equivalent to minimizingjwj 2 . Therefore, the optimization problem, which is an example of a quadratic programming problem, that should be solved is: [98]: argmin w;b 1 1 jwj 2 (3.25) 58 In case data are not linearly separable, exact separation of the training data can lead to poor generalization. Therefore, the model needs to be modified to allow the misclassification of some data points with a penalty that increases with that boundary’s distance. Hence, the slack variables, x n > 0 where n= 1;:::;N, with one slack variable for each training data point, is introduced [99], where x n = 8 > > < > > : 0 Data points on or inside the correct margin boundary jt n y(x n )j otherwise (3.26) Thus, for data points that are correctly classified and are either on the margin or on the correct side of the marginx n = 0, and for points inside the margin 0<x n < 1. For the misclassified points and those on the wrong side of the decision boundaryx n > 1. It relaxes the hard margin constraint to give a soft margin and allows misclassification of some data points. The problem now is to maximize the margin while softly penalizing points on the wrong side of the boundary, which is equivalent to minimizing function defined as: C N å n=1 x n + 1 2 jwj 2 (3.27) where the parameter C> 0 controls the trade-off between the slack variable penalty and the margin [98]. Decision Trees Decision trees are among the most popular and powerful machine learning algorithms that analyze data to construct a set of rules for predicting labels. They can work with both categorical and nu- merical data. A common learning strategy employed by different algorithms for constructing a tree from data is an iterative process of splitting the data into subsets, known as recursive partitioning until the stopping criterion is met. This process is an example of a greedy algorithm that uses a set of rules to make locally optimum decisions on selecting the best tributes for data partitioning. Sev- 59 eral popular decision tree induction algorithms, such as classification and regression trees (CART), ID3, and C4.5, are working based on this method. The gain is a criterion that can determine the quality of a split. It compares the impurity degree of the child nodes (after splitting) with the parent node (before splitting) and defined as [101]: Gain= I(parent) k å i=1 N(v i ) N I(v i ) (3.28) where I(:) is the impurity measure of a given node ,N is the total number of items at the parent node, k is the number of attribute values, and N(v i ) is the number of items associated with the child node v i . Different algorithms employ various metrics to measure the split’s quality. Entropy, Gini impu- rity, and classification error are examples of impurity measures. The entropy of a random variable from information theory represents its average amount of information. It the basis of algorithms such as ID3 and C4.5. Another metric is Gini impurity, which is used by CART. It represents the probability of incorrectly labeling a randomly selected data point from the set if it were randomly classified according to the class distribution in the dataset. These measures are defined as follows [98, 101]: Gini(n)= 1 C å i=1 p(ijn) 2 (3.29) Entropy(n)= C å i=1 p(ijn) log p(ijn) (3.30) Classification Error(n)= 1 max i p(ijn) (3.31) where p(ijn) denotes the probability of randomly picking an element of class i at a given node n, and C is the number of classes. 60 Random Forest Random Forest is one of the most powerful and popular machine learning algorithms that employs bootstrap aggregating (bagging). It constructs multiple decision trees at training time and repeat- edly trains each decision tree with randomly selected subsets of training data. By aggregating the predictions from all decision trees, the label of test data can be decided. 3.5.2.3 Recurrent Neural Network (RNN) An RNN is a special class of artificial neural network models, which contains cyclic connections [102]. Unlike the feedforward neural networks, RNNs are able to learn complex temporal dynamic sequential data. An RNN cell maps the input x t to hidden states h t and output y t as follows [103]: h t =j h (W h h t1 + U h x t + b h ) (3.32) y t =j o (W y h t + b y ) (3.33) where W ;U are the weight matrices, b is the bias vector, andj h (:) andj o (:) are element-wise, non-linear activation functions (e.g., hyperbolic tangent, rectified linear unit (ReLU), and sigmoid) in the hidden layer and the output layer. Although the RNNs have shown great success in many sequence labeling and detection ap- plications, their training can be challenging because of vanishing or exploding gradient problems. The LSTM-based RNN provides a solution to overcome these modeling weaknesses and to learn long-term dependencies [104, 105]. A common LSTM cell comprises input gate i t , output gate o t , forget gate f t , and memory cell c t . The gates allow the network to learn when to forget or update the hidden states h t given the new information, and the current output y t is considered equal to the current hidden state. The function of each cell is derived as follows [103, 106, 107]: i t =s(W i h t1 + U i x t + b i ) (3.34) f t =s(W f h t1 + U f x t + b f ) (3.35) 61 o t =s(W o h t1 + U o x t + b o ) (3.36) c t = f t c t1 + i t tanh(W c h t1 + U c x t + b c ) (3.37) h t = o t tanh(c t ) (3.38) wheres(:) denotes the sigmoid function, and is the element-wise product. A recurrent network can have multiple nonlinear hidden layers of RNN cells between input and output layers, which is considered a deep recurrent neural network (DRNN). Using deep architectures can result in higher data representations, and a more accurate model [103]. Assume that the network has L hidden layers, and the output sequence of a layer is the input sequence for the next layer. Using equations (3.32)-(3.38), the output y l t , hidden states h l t , and memory cell c l t can be computed for layer l from 1 to L as follows [106, 108]: [y l t ;h l t ]= VanillaRNN(y l1 t ; h l1 t ; q l ) (3.39) [y l t ;h l t ;c l t ]= LSTM(y l1 t ; h l1 t ; c l1 t ; q l ) (3.40) whereq l represents the parameters (W l ; U l ; b l ) of the RNN cells for layer l, and y 0 t = x t . A schematic diagram of the neural network structure is summarized in Figure 3.9. The set of magnetic induction (MI) signals observed by the coils at time t is considered as the input vector x t . A time window of 1 second (T = 1s) is sliding over the data with 0.5 second overlap and feeding the truncated subsequences of input data within the window to the batch normalization layer. Then the normalized input data(x tT+1 ;:::; x t1 ; x t ) is fetched to the deep long short-term memory (LSTM) model. The network outputs sequences of vectors(y L tT+1 ;:::; y L t1 ; y L t ), where each output vector shows the prediction score of its corresponding input sample. Assuming the input signals are sequenced to N samples, the overall score of the entire window can be calculated by averaging all of the scores within the window into a single prediction vector of scores ˆ y t [106]. Then the prediction scores are converted into class membership probabilities b O t by applying a 62 Figure 3.9: Architecture of deep recurrent neural network (RNN). softmax layer. The predicted class membership probability vector contains the probability of every class generated by our model. The most probable class is then selected as the predicted activity label for the given input data within the time window. The deep LSTM model is implemented in the TensorFlow framework. We used the mean cross- entropy between the ground truth labels and the predicted class membership probability vector as the loss function, and the network parameters are updated by minimizing this loss function. The model is trained using batch gradient descent with the RMSprop updating rule. The entire training set is passed through the neural network model to update the model with an exponentially decaying learning rate in each training epoch. The dropout regularization technique is also applied to all nodes in the network to avoid overfitting. The dropout keep-probability determines the probability of keeping a node during training. After each epoch, the performance of the model is evaluated on the validation set. We evaluated the influence of several hyperparameters related to the network architecture and learning process using grid-search. These hyperparameters and their range of values explored for tuning during training are summarized in Table 3.2. We implemented a 5-layer network with 20 and 40 units for BML and MHAD datasets, respectively. Both datasets are trained with the optimizer decay rate of 0.95, the initial learning rate of 0.01, the exponential decay rate of 0.98, exponential decay step of 100, and the keep probability of 0.8. 63 Table 3.2: Hyperparameters related to the deep recurrent neural network architecture and learning and their range of values explored for tuning during training. Hyperparameter Range Number of Layers f1, 2 ,3, 5, 10g Number of Units f5, 10, 15, 20, 30, 40, 50g Keep Probability f0.2, 0.5, 0.8, 1g Optimizer Decay Rate f0.8, 0.85, 0.9, 0.95, 0.98g Optimizer Momentum f0.001, 0.01, 0.1g Initial Learning Rate f0.001, 0.01, 0.1g Exponential Decay Rate f0.85, 0.9, 0.95, 0.98g Exponential Decay Step f50, 100, 200, 300g 3.5.3 Results Tracking the motion of body parts during physical activity is critical in characterizing an indi- vidual’s movement, and collecting data that provide a more accurate representation of these mo- tions results in better activity detection. The MI signals express a strong relationship with the geo-translation of body segments since the system gain is directly affected by distance and mis- alignment between coils. Distributing more coils around the human body provides comprehensive information about the user’s body movements and results in a better distinction between similar actions. We used the MHAD dataset to compare the capability of the MI signal and accelerom- eter data in estimating the location of a body part during physical activity. The accelerometer is considered here as a benchmark because it is the most frequently used wearable sensor modality for human activity monitoring. Six markers placed close to the accelerometers are considered as target points. Then the similarity between the 3D location of each target point and data of its corresponding accelerometer and MI transceiver is calculated. We used R-squared (R 2 ) as the similarity metric, and the average values over the whole dataset are presented in Figure 3.10. The R-squared reports the similarity between two sets of data by a number between zero and one, where a higher number shows a stronger relationship between two 64 Figure 3.10: Average R-squared between XYZ of each target point and data of its corresponding accelerometer and magnetic induction (MI) transceiver. datasets. The results show that, on average, the MI signal has a stronger relationship with the 3D location of markers compared to the accelerometer data. This characteristic can be useful not only in classifying human activities but also in reconstructing the motion trajectories of body segments. Many studies have adopted IMUs to reconstruct the trajectories of movements for motion analysis in different applications. Examples include handwritten digit recognition [109], monitoring trunk kinematics during standing up to sitting down [110], and tracking the motion of body parts on patients who have been affected by neurological conditions for rehabilitation purposes [111]. In inertial sensor-based recognition systems, the velocity and positions are computed indirectly by the integration over sensor measurements. It makes the estimation errors caused by the intrinsic noise/drift grow un- bounded with time. For example, the average displacement error of Xsens IMU after one minute is about 152 meters [111]. On the other hand, the MI motion signal is directly affected by the location and orientation of coils. As a result, the trajectory reconstruction using MI signals does not require integration over measured data, which removes the problem of the cumulative error. Table 3.3 summarizes the performance results of LSTM with methods including SVM, KNN, DT, RF, and LR. The confusion matrix of each classification method on BML and MHAD datasets are also presented in Figure 3.11 and Figure 3.12, respectively. The rows and columns represent the percentage of true activity labels and the predicted activity labels, respectively. The results are compared to other previously introduced methods using different modalities for activity detection. 65 Table 3.3: The performance summary of classification results using generated synthetic magnetic induction (MI) motion data of different datasets. Classifier Model Overall Accuracy (%) Average Precision (%) Average Recall (%) F1 Score BML SVM 83.5 85.4 83.5 0.84 KNN 79.4 80.0 94.4 0.8 Decision Trees 77.3 77.1 77.1 0.77 Random Forest 86.6 86.5 86.5 0.86 Logistic Regression 83.5 86.2 83.6 0.85 Deep LSTM 87.0 86.7 87.0 0.87 MHAD SVM 96.4 96.6 96.4 0.96 KNN 90.3 91.1 90.3 0.91 Decision Trees 81.2 82.3 81.2 0.82 Random Forest 90.9 91.8 90.9 0.91 Logistic Regression 90.9 91.3 90.9 0.91 Deep LSTM 98.9 98.9 98.9 0.99 Table 3.4: Performance comparison with other state-of-the-art methods using different modalities for activity detection. Multi-Task Conditional Restricted Boltzmann Machines (MT-CRBMs). Classifier Modality Accuracy (%) Reference BML SVM Motion Capture 41.3 [112] MT-CRBMs-Deep Motion Capture 54.5 [112] MHAD SVM Accelerometer 98.0 [95] Random Forest Motion Capture 96.0 [113] Random Forest Audio 68.2 [113] LSTM Camera (RGB Image) 92.4 [114] 66 (a) SVM (b) KNN (c) DT (d) RF (e) LR (f) Depp LSTM Figure 3.11: Confusion matrix for the validation set corresponding to the Biological Motion Li- brary (BML). We employed accuracy as an evaluation metric for comparison, as datasets used here are balanced and have an equal number of samples for each activity. The results are also compared to other previously introduced methods for activity detection, and the reported results on these datasets are summarized in Table 3.4. Our results indicate that the deep LSTM model with optimum hyperparameters outperforms other classifiers by a considerable margin on the generated synthetic MI motion data. The recurrent neural networks can capture sequential and time dependencies between input data that results in a strong performance. The LSTM cells let the model capture even longer dependencies compared to vanilla cells. A deep architecture with an optimal number of layers enables the neural network to extract useful discriminative features from the set of input data and to improve the performance of the model. It should be noted that the datasets used here are diverse, which proves the classifier models are valid for a broad range of activity recognition tasks. Moreover, the actions recorded in the BML dataset, including knocking, lifting, and throwing, are very similar as only one hand is moving. 67 (a) SVM (b) KNN (c) DT (d) RF (e) LR (f) LSTM Figure 3.12: Confusion matrix for the validation set corresponding to the Berkeley Multimodal Human Action Database (MHAD). 68 The same movement of human body parts in these activities makes it difficult to distinguish and categorize them. Despite these challenges, the deep LSTM model has achieved high accuracy, and it indicates that the recurrent model is capable of classifying human actions by using MI motion signals. 3.6 Conclusion Human activity recognition is a powerful technology with a wide range of applications such as healthcare, rehabilitation, sports training, and senior monitoring. We proposed a new wearable- based HAR system using magnetic induction for motion capture and wireless signal transmission. This method can tackle existing issues with conventional HAR systems in various aspects, includ- ing power consumption, the complexity of implementation, and cost. It can also provide a suitable infrastructure for new applications working in harsh environments, such as underwater. The pro- posed system is a new sensing approach for capturing human motions, which can also be integrated with other monitoring modalities to provide a more comprehensive HAR system. To show the capability of the MI-HAR system in detecting human movements, we generated synthetic MI motion data received from MI transmitters around the user’s body during different activities by the MI system model. As mentioned before, the model used for synthesizing MI mo- tion data does not consider cross-coupling between transmitter coils. However, this cross-coupling is not necessarily destructive and can even provide further information regarding the location and alignment of all coils relative to each other. In this scenario, each received signal is not only a function of the transmitter and receiver coils but also the arrangement of all other coils affects it. Therefore, the movement of even a single body part results in a different signal pattern and can make the system more accurate in detecting actions similar to each other. We employed several commonly used machine leaning-based classifiers and deep recurrent neural networks for the detection step. We empirically evaluated the proposed MI-HAR system by conducting experiments on the generated synthetic MI motion dataset and discussed the outcomes in detail. Experimental results reveal that the proposed deep LSTM model shows outstanding 69 performance compared to other approaches. One of the benefits of using the deep recurrent neural network for sequence classification is that it can support multiple parallel temporal input data from different sensor modalities such as MI sensors, accelerometers, and gyroscopes. The model can learn complex features directly from raw data and map them to activities. It removes the need for manual feature engineering by experts while it achieves a comparable per- formance to models with the feature handcrafting step. Besides, the neural network model enables an interactive learning system when the user provides training data even after the initial training step. It allows the user to fine-tune a pre-trained neural network model with their personal data. However, the neural network complexity should be assessed where models have to be implemented in embedded systems with limited processing capability. It highlights the importance of trade-off between computational cost and detection accuracy to ensure real-time feedback. 70 Chapter 4 3D Motion Tracking 4.1 Introduction Over the past decade, monitoring and recognition of human activities have embraced a growing number of practical usage in a broad range of domains such as healthcare, rehabilitation, sports training, virtual reality (VR) gaming, human-computer interface (HCI) systems, finger tracking, daily life-logging, child and elderly care, and assistance for people with cognitive disorders or chronic conditions [77–79]. Tracking and reconstructing limb movements in 3D space facilitates more detailed evaluation, and it is crucial for the analysis and clinical understanding of complex functional movements. Studying biomechanics of human motion has application human perfor- mance assessment, gesture/posture monitoring, behavioral recognition, gait analysis, and patients’ functionality and improvement evaluation during the rehabilitation period [115–118]. There are many solutions for tracking body movement using different monitoring sources [119]. Computer vision-based methods, such as Kinect or optical MoCap system, are the most commonly used techniques that allow users to interact with them and collect data on the user’s motion using depth sensors, color, and infrared cameras [120, 121]. However, they inherit com- puter vision limitations such as light dependency, coverage limitation, and high computational cost [82, 122]. The MoCap systems require an expensive setup of infrared cameras for tracking reflective markers on an individual’s body, which makes them only applicable to the laboratory en- 71 vironment and restricted in physical space. Besides, the markers placement and soft tissue artifacts have a considerable effect on the system accuracy [81, 123]. The RF-based solutions are another motion tracking method capturing data based on wireless signal changes (e.g., Doppler frequency shift and signal amplitude fluctuation) [124]. These methods also suffer from environmental de- pendency and limitation in the number of detectable gestures due to the high cost of training data collection and the lack of multi-user identification capabilities [125]. Wearable-based solutions are an alternative, cost-effective solution for applications where the optical-based methods are unsuitable. This approach tracks the user’s movement based on the sensors readings placed around the human body [126]. The advancement of sensing technolo- gies, miniaturization, embedded systems, and wireless communication systems combined with predictive models for data analysis and detection have made it possible to develop wearable de- vices working around the human body for continuous physical activity monitoring. Smart devices like smartphones, smartwatches, and fitness bands are becoming widespread for providing use- ful insights about an individual’s performance and health status. These wearables have multiple embedded physiological, inertial, and ambient sensors that enable multi-modal sensing [42]. Many studies have exploited commercial IMU devices comprised of accelerometers, gyro- scopes, and magnetic sensors, for motion tracking based on wearable sensors. An IMU can be attached to a body segment to estimate its movement in space. By combining multiple of them on adjacent body segments, the kinematics of activities can be determined [126]. For example, [127] presents the development of a smart wearable jumpsuit with multi built-in IMU sensors for automatic posture and movement tracking of infants. The study presented in [128] investigates the reliability and validity of IMUs for clinical movement analysis, and [129] presents a single wrist- worn IMU sensor for high-resolution motor state detection in Parkinson’s disease. Inertial sensing can track limb movements by integrating over sensor measurements, though it is subject to drift since the estimation errors caused by the intrinsic noise can grow unbounded with time [42]. 72 In some applications, it is possible to achieve improved accuracy and more specific inferences by fusing the subsets of data collected from those sensors compared to single sensor modalities [130, 131]. Although these devices provide a solution for physiological health monitoring, condi- tion assessment, and medical diagnosis, they still might face challenges. These single-node devices restrict biosensors’ placement while optimizing their position can increase the system’s accuracy and robustness in monitoring vital signs (e.g., body temperature and heart rate) [18, 42]. Fur- thermore, in systems relying on data from a single device, variations in position can significantly affect the performance. In motion tracking applications, a single node wearable is not able to cover the entire body. Therefore it cannot get detailed information about the mobility of an individual’s limbs. For example, a smartwatch’s inertial sensor cannot capture the movement of the user’s legs, limiting the system’s ability in classifying activities [42]. A network of distributed wearable devices operating around the human body is an approach that can address these issues. One of the biggest challenges in wearable-based motion tracking systems is to find the optimum type and number of non-invasive sensors with minimal power consumption to achieve acceptable accuracy and satisfy guidelines and constraints. In [42], we introduced a wireless system based on magnetic induction combined with machine learning techniques to detect a wide range of human activities. We showed that this system can address challenges in terms of power consumption, accuracy, coverage, privacy, and cost. We investigate the capability of the MI system in 3D motion tracking and evaluate a prototype device in estimating its motion using trained machine learning- based regressors. We use a calibrated MI model to generate synthetic MI motion data and train regressors without the need for any measured data. This chapter is organized as follows: in Section 4.2, the system principle, architecture, and hardware development are presented. In Section 4.3, data collection procedures, including motion and MI sensor data, are explained. Generating synthetic data using the circuit model of an MI system is also described in this section. The tracking algorithm, system evaluation, and experi- 73 mental results are presented in Section 4.4. Section 4.6 presents future works and discusses design challenges in developing a practical sensing system from the presented prototype, and Section 4.7 concludes the chapter. 4.2 System Design 4.2.1 Operating Principles As explained in chapter 2, the MI-based communication system is a short-range wireless physical layer that transmits signals by inductive coupling between the wire coils rather than radiating as conventional methods [34, 55]. The transmitter node uses a coil to produce an oscillating magnetic field at a specific frequency. Each sensor node’s (receiver) main component is a coil to capture the transmitter’s generated magnetic field. According to Faraday’s law, the time-varying magnetic field induces a voltage in sensor nodes proportional to the rate of magnetic flux change through their coils. For a predefined coil geometry and operating frequency below 30 MHz, where the environmental effects are negligible, the flux change rate is a function of the sensor coils’ position and orientation relative to the transmitter [34, 42]. The relationship function from spatial data into induced voltage is non-linear and surjective, and the tracking problem objective is to estimate the sensors’ position given the induced voltage measurements. 4.2.2 Architecture We used an analytical model of the MI system presented in [34, 42] to calculate the induced voltage at each sensor coil given its position and orientation. It forms the basis of the data-driven backward estimation algorithm that retrieves a node’s position using its observed data. It helps assess the system performance under different configurations, such as changing the number or arrangement of sensor coils to find the near-optimal setup with acceptable tracking accuracy. Since the model is a function of relative distance and alignment of coils to the transmitter, we transform the coordinate system to locate the new coordinate system’s origin at the center of the transmitter coil, with the 74 (a) Single sensor setting (b) Orthogonal setting (c) Parallel setting Figure 4.1: Data-driven location tracking across all settings for a predefined target point. coil’s surface normal oriented in the z-direction. Given the sensors’ spatial data, we compute the coordinate transformation matrix and calculate the coil’s position and orientation in the new coordinate frame. We explored the node’s position p=(x;y;z) with the resolution of 1 cm, and alignment of: ˆ n=(sinq cosf; sinq sinf; cosq) (4.1) With the resolution of 5 in space domain given as: x2[20 cm; 20 cm] (4.2) y2[20 cm; 20 cm] (4.3) z2[60 cm;10 cm] (4.4) q2[0 ; 60 ] (4.5) f2[0 ; 360 ] (4.6) For a given set of observed data, the possible solutions are retrieved, which is a single point solution in an optimal configuration. 75 We studied the performance of an MI sensor (single sensor setting), where the coil can be aligned in any direction. We also adopted two-sensor configurations and investigated different alignment setups. Among these setups, we present the performance analysis of setups where coils’ surface normal are aligned in the same direction (parallel setting) or perpendicular to each other (orthogonal setting). In these experiments, the induced voltage measured at the coils is used as input for location estimation. Figure 4.1 shows an example result of the data-driven backward estimation algorithm, where each point represents a possible node’s position with at least one alignment that can produce the given set of inputs for the defined setting. As the results display, there are many possible solutions for a single sensor setup, and this number reduces by adding another sensor. The sensor data are assumed to be measured with 1 mv accuracy and given as inputs to the algorithm. A comparison between two-sensors configurations shows that the parallel setting outperforms the orthogonal setting. Although a unique solution is not returned as an output, results suggest that the regression methods with proper constraints can meet the minimum required accuracy for the position tracking. 4.2.3 Hardware The system consists of a transmitter (central) node generating an oscillating signal at 13.56 MHz. We used ISC.LRM1002 long-range RFID reader module [132] attached to ISC.ANT310/310 long- range HF antenna [132] to generate RF signal. Since we used this setup for RFID measurements presented in the Section 4.6, we used the same transmitter for a better comparison. The receiver node consists of MI sensors. Each sensor includes an air-cored, single-layer copper coil with a 5 cm radius and 10 AWG wire diameter to capture the transmitter’s signal and measure the induced voltage. Resistance and self-inductance of the coil measured by VNA at the resonance frequency are 0.101 W, and 241 nH, respectively. To improve the system efficiency, we have employed resonant inductive coupling attached to the coil. The tuning circuit can be as simple as a capacitor to tune the frequency or be a P or T matching circuit to tune the frequency, control 76 Figure 4.2: The MI sensor prototype hardware contains 1) variable capacitor for frequency tuning, 2) an envelope detector, and 3) Arduino microcontroller for measurement. Q-factor, and match input and output impedances for higher power transfer. Here, we used a 560 pF capacitor parallel to a trimmable capacitor with the adjustable range of 3-10 pF to tune the circuit to resonance accurately. The transmitted AC signal attenuates as a function of distance and alignment of the node with respect to the transmitter antenna. To track the signal’s amplitude changes, we used an envelope detector consisting of an IN5817 Schottky diode, a resistor or 1 KW, and a capacitor of 1 nF. The envelope detector’s output, which is the resistor’s voltage, is measured by an Arduino Nano (ATmega168) microcontroller. The resolution of ADC (analog pin A1) is 10 bit for a defined measurement range. Figure 4.2 depicts an MI sensor components. 4.3 Data Collection 4.3.1 Measurements We employed a Microsoft Kinect v2 to capture the 3D position and alignment of the transmitter and the MI sensor node. The Kinect sensor consists of a depth camera, an RGB camera, and a microphone array sensor. The RGB camera and depth camera respectively provide a 19201080 color image, and 512424 depth image at 30 frames per second with a resolution of a few mil- limeters in measure range between 0.5 m to 4.5 m [133]. The depth stream provides the sensor’s 77 distance to every point within its area of coverage. As the cameras have different pixel resolutions and are not perfectly aligned, three coordinate spaces and types are defined: color space point (x c ;y c ), depth space point(x d ;y d ), and camera space point(x w ;y w ;z w ), representing a point in the color images, depth images, and real-world, respectively. The software development kit (SDK)’s mapping function can be used to map a point from one coordinate space to another. We used colored markers to facilitate motion tracking of the devices and developed a video processing algorithm analyzing the color frames to locate pixels corresponding to the target color. The transmitter antenna and the MI node are labeled with distinct colored markers and placed in front of a white background. A threshold range is set for each color to extract pixels with the color value within the defined range. The detected pixels are classified to N m clusters, where N m is the number of markers, using K-means clustering methods. Then, the connected neighboring pixels of each cluster are grouped. Since the markers are colored foam balls, the circle with the minimum area enclosing each set is calculated, and the largest region is given as the target circle. The next step is mapping color to camera space to find the corresponding spatial location of each extracted color pixel. The result is a list of 3D real-world points mapped from the target circle’s pixels, and each marker’s location is computed by taking median over all the calculated values. This process repeats for each new color frame that Kinect captures. The analytical model requires the center and alignment of the transmitter and receiver coils/antennas as inputs to estimate the induced voltage. For determine a coil’s surface normal, at least three markers (M i : i2f1;:::;N m g with N m >= 2) are required. Hence, we used four red and three blue markers to track the transmitter antenna and the MI sensor node. The center of each device is calculated by averaging over its markers’ location c=å N m i=1 M i , and its surface normal is also calculated by the cross product of vectors passing through the markers as follows: ˆ n=~ v 1 ~ v 2 (4.7) where ~ v 1 =(M 1 M 2 ) (4.8) 78 Figure 4.3: Schematic representation of measurement setup. ~ v 2 =(M 1 M 3 ) (4.9) We applied the median filter, a non-linear digital filtering technique, to remove noise and spikes in the extracted location and alignment data. The induced voltage, V ind , at the MI sensors is measured for 30 seconds via Arduino by using a Python script that controls the recording in order to synchronize Kinect’s motion data and Ar- duino’s measurements. The sampling frequency is 100 Hz, and the reference voltage range is 0 V to 5 V , which results in the quantization interval of 5/1024 V . The data streams of the node’s MI sensors are recorded and used as inputs for the regression model to estimate the device’s location. The sampling rate of motion data recorded by Kinect and the sensors’ data are different. Therefore, all recordings are resampled with a sampling interval of 100 ms, which also handles the missing sample values. The measurement setup of experimental measurements is presented in Figure 4.3. 79 4.3.2 Synthetic Data Validating and testing a machine learning model is a critical stage in model development. It is challenging because of the difficulty of collecting realistic and valid data and the lack of labeled data. One solution is to create synthetic data for training the model, and here, we used a Varia- tional Auto-Encoder (V AE) model to produce time-series motion data. An V AE is based on the auto-encoder architecture and is composed of encoder and decoder networks. The encoder com- presses the data into a lower-dimensional space called the latent space representation. The decoder decompresses the reduced representation code to reconstruct the original data. The V AE learns the probabilistic interpretation of these networks and generates new samples using different latent variables as input. Consider datasetfx (i) g N i=1 is consists of N i.i.d. samples of some variable x. V AEs assume that the data are generated by a random process with continuous latent variable, and each latent variable z is related to its corresponding observation x through likelihood p q (xjz), where p q is a probability distribution with parametersq. This probabilistic interpretation of the decoder can decode a latent (hidden) representation code into a distribution over the observation. Similarly, the encoder net- work returns a latent code sampled from the posterior density distribution p q (zjx) given a sample from data space [134]. While both prior p(z) and likelihood p(xjz) can be formulated exactly, the posterior p(zjx) requires an intractable integral over the latent space. Hence, an approximate posterior q f (zjx) closest in Kullback-Leibler (KL) divergence to the actual, intractable posterior distribution is considered. The approximate posterior is parameterized by variational parameters f, and the training objective is a tractable lower bound to the log-likelihood [135]: log p(x) E q f (zjx) h log p q (x;z) q f (zjx) i =L(x;q;f) (4.10) and can be equivalently written as: L(x;q;f)=E q f (zjx) [ log p q (xjz)]D KL (q f (zjx)jjP q (z)) (4.11) 80 On the right-hand side of equation (4.11), the first term, reconstruction error, represents the likelihood of the model reconstructing the input data. The second term, variational regularization term, is the Kullback-Leibler divergence and makes the approximate posterior q f (zjx) to be close to p q (z. TheL(x;q;f) is a lower bound on the log probability of data p q (x), which is called evidence lower bound (ELBO). Maximizing ELBO with respect to the model parameters q and variational parametersf respectively maximizes the marginal probability p q (x) and minimizes the KL divergence [134]. We trained the model using the sensors motion data tracked by the Kinect to produce synthetic time-series samples. After training the model, new time-series data can be generated by sampling from latent space z with normal distribution parametrized by the mean and the variance [135]. The generated data includes the motion of the coils’ center and alignment in 3D space for a predefined sensor setting. We synthesized angular variables q and f to calculated the corresponding coil’s surface normal ˆ n that can be defined as: ˆ n=(sinq cosf; sinq sinf; cosq) (4.12) where the variablesq andf can take values in the range of 0-90 and 0-360 degrees, respectively. To generate training data for the motion tracking algorithm, we synthesized the induced voltage at the MI sensor using the two-port network model of the MI system [34, 42]. The circuit model represents a forward model of the system estimating the sensor measurements given node motion data. To evaluate the circuit model’s accuracy in synthesizing MI data, we fetched the captured motion data by the Kinect as inputs and estimated the corresponding induced voltage at the sensors. The data streams corresponding to MI sensors are simulated for each motion sample. The circuit model is calibrated by finding the scale and bias of synthesized data with respect to the measure- ments. Considering s i and m i as the generated synthetic data and measurements corresponding to 81 Figure 4.4: The measured and simulated induced voltage at an MI sensor during two arbitrary movements, such that both relative alignment and location of the coil varies. a motion sample, the scale a and bias b can be calculated as follows: a= 1 N s N s å i=1 s si s mi (4.13) b= 1 N s N s å i=1 m si s si s mi m mi (4.14) where m si ;s si ;m mi ;s mi represent the mean and standard deviation of synthetic data and measure- ments corresponding to motion sample ith from N s samples. We have performed the experiment for 220 motions, including different spatial translation and rotation (N s =220). Figure 4.4 shows the measured and simulated data of sensors during their movement, picked from the evaluation dataset after calibrating the model. The average NRMSE and cross-correlation of the synthesized and measured data for all experiments are 12% and 0.91, respectively. It should be noted that the reported metrics consider not only the MI system model 82 inaccuracy but also the error associated with the Kinect-based marker tracking algorithm and Ar- duino measurements. The measured motion data samples are also used for training V AE to generate synthetic motion data. 4.4 Evaluation 4.4.1 Machine Learning-based Regression Regression analysis is a set of machine learning methods for estimating a continuous outcome (dependent) variable y based on the value of predictor (independent) variables by defining a math- ematical relationship. It has many real-world applications, such as predicting stock market price given current market conditions, location tracking in 3d space, weather forecasting using weather history, sensor data, etc.[99]. We implemented machine learning regression algorithms to solve the reverse problem of esti- mating a node’s 3D position (x,y,z) from its sensors’ measurements in meters. The performance of several regression models, including extremely randomized trees or extra trees (ET), RF, KNN, light gradient boosting machine (LightGBM), multi-layer perceptrons (MLP), DT, and linear regression (LR) is compared using PyCaret [136], an open-source machine learning library in Python. A short description of these models is provided here. Linear Regression Linear regression is a widely used method that estimates the output variable as a linear combination of the inputs by the scalar product between the input vector x, the model’s weight vector w, and bias parameter b: y(x;w)= w T x+ b (4.15) 83 The model can be extended by using nonlinear basis functions f. It allows the function to be a nonlinear function of the input vectors. y(x;w)= w T (x)+ b (4.16) K-Nearest Neighbors KNN regressor employs the same technique and distance functions as its classifier. Instead of using the neighbors’ labels, it takes their value and averages them to estimate the value of new test data. Decision Trees To use a decision tree for regression, an impurity metric suitable for continuous variables is re- quired. For example, mean-squared error (MSE) and mean absolute error (MAE) are impurity measures that can be used for the regression problems, which respectively minimize the L2 and L1 losses. Random Forest The bagging learning algorithm repeatedly (N times) selects, with replacement, random samples and their labels from the training set to train a decision tree T i . The target value for a new data sample x can be calculated by averaging the predictions from all the regression trees on this new sample [99]: y= 1 N N å i=1 T i (x) (4.17) Extremely Randomized Trees (Extra Trees) It is an ensemble learning method similar to the random forest model that outputs the target value by aggregating multiple decision tree results. The ET method differs from the RF in building and training the decision trees. The ET method uses the whole dataset, whereas the RF subsamples the 84 input data with replacement. Additionally, the splitting is randomized in ET instead of selecting the optimum split based on, for example, Gini impurity or entropy. A random value in the feature range of each feature under consideration is selected, and then the best is selected as the split [137]. Light Gradient Boosting Machine The LightGBM is a fast distributed gradient boosting framework based on the decision tree algo- rithms developed to improve performance and scalability. It selects the leaf returning the largest decrease in loss and splits the tree leaf-wise, whereas level-wise in other boosting algorithms. This approach is much faster than the level-wise algorithm and has better success in loss reduction, and therefore, achieves considerably higher accuracy. Leaf-wise splits increase complexity and may lead to overfitting, which can be dealt with by specifying the max depth that the tree is allowed to grow [138]. Multi-layer Perceptron An MLP is a feed-forward ANN consisting of neurons called perceptrons. A perceptron’s output is calculated from a linear combination of weighted inputs. The output can also pass through a nonlinear activation function, and therefore the final result can be written as [139]: y=s(W T x) (4.18) where x is the input vector, w is the weight vector, y is the output, and s denotes the activation function. An MLP network consists of at least three layers, including an input, an output, and one or more hidden layers. MLPs are fully connected, which means that each perceptron is connected with a certain weight to all neurons in the next layer. Except for the input layer, perceptron in other layers use nonlinear activation functions, such as logistic and hyperbolic tangent. It allows the network to model complex and nonlinear relationships between inputs and targets and therefore provides a richer capability in distinguishing data that are not linearly separable [140]. For training, MLPs 85 Figure 4.5: The evaluation process, including training with synthetic data, evaluation on synthetic data, and test on the real-word measurements. utilize supervised backpropagation learning technique to minimize the cost function by iteratively adjusting the weights. The gradients of the cost function determine the level of adjustment with respect to parameters. 4.4.2 Results The regression models are trained on 70% of synthetic data and then scored on the remaining data using 10-fold cross-validation method. The metrics used for comparison are RMSE, mean absolute percentage error (MAPE), and R 2 . Before fetching data into the regressors, each feature is standardized individually, and the missing values are substituted with previous non-missing values. The processed data are then divided into fixed-length segments of 2 seconds using the sliding window technique with a 0.1 second step size. Table 4.1 summarizes the performance results of all models on the synthetic data for different settings. As the results show, the moving node’s distance and position in the z-direction with respect to the transmitter coordinate frame can be tracked with acceptable accuracy. It means that a system with an additional transmitter in a different direction can track the node’s motion in the new direction, enabling 3D positional tracking. To provide a realistic assessment of real-world performance, we evaluated each of the optimal models’ tracking accuracy on measured data as well. The evaluation process is depicted in Figure 4.5. According to the score measures reported on synthetic data, the LightGBM regressor in the single sensor setting and the extra trees regressor in two-sensors (orthogonal and parallel) settings outperform other models. Figure 4.6 presents the 86 Table 4.1: Performance of regression models in motion tracking using synthetic data generated for the different settings. Model Distance X Y Z RMSE MAPE R2 RMSE MAPE R2 RMSE MAPE R2 RMSE MAPE R2 Single Sensor ET 0.028 0.043 0.898 0.05 6.548 0.232 0.035 3.877 0.102 0.031 0.05 0.868 RF 0.028 0.042 0.9 0.05 6.872 0.25 0.035 3.88 0.125 0.031 0.049 0.871 KNN 0.029 0.044 0.89 0.052 6.155 0.161 0.037 3.923 0.017 0.033 0.052 0.856 LightGBM 0.028 0.042 0.9 0.048 6.873 0.29 0.034 3.931 0.173 0.031 0.049 0.872 MLP 0.03 0.045 0.883 0.049 6.806 0.278 0.034 3.999 0.147 0.034 0.053 0.846 DT 0.039 0.057 0.807 0.069 10.529 -0.45 0.049 4.05 -0.708 0.043 0.067 0.753 LR 0.04 0.066 0.796 0.052 8.079 0.177 0.035 4.095 0.138 0.041 0.069 0.769 Orthogonal ET 0.022 0.029 0.879 0.04 4.897 0.537 0.011 0.851 0.18 0.005 0.007 0.169 RF 0.023 0.031 0.863 0.043 5.266 0.47 0.031 2.53 0.362 0.026 0.037 0.827 KNN 0.022 0.029 0.875 0.042 5.749 0.487 0.03 2.137 0.387 0.025 0.034 0.84 LightGBM 0.026 0.036 0.826 0.05 5.989 0.294 0.035 3.018 0.158 0.029 0.043 0.779 MLP 0.026 0.037 0.823 0.05 5.239 0.277 0.036 3.191 0.134 0.03 0.043 0.774 DT 0.033 0.041 0.718 0.062 7.04 -0.096 0.044 2.71 -0.313 0.037 0.049 0.642 LR 0.035 0.05 0.688 0.053 6.829 0.184 0.037 3.2 0.103 0.038 0.057 0.625 Parallel ET 0.011 0.018 0.947 0.037 7.539 0.365 0.028 5.184 0.23 0.014 0.025 0.917 RF 0.011 0.019 0.94 0.037 9.49 0.334 0.028 5.471 0.193 0.014 0.026 0.908 KNN 0.012 0.021 0.932 0.037 10.543 0.352 0.028 4.855 0.196 0.015 0.027 0.9 LightGBM 0.012 0.021 0.932 0.04 10.004 0.261 0.03 5.827 0.075 0.015 0.028 0.897 MLP 0.013 0.022 0.924 0.04 9.251 0.231 0.031 6.268 0.022 0.016 0.03 0.881 DT 0.016 0.026 0.882 0.053 14.302 -0.338 0.04 5.478 -0.636 0.02 0.035 0.818 LR 0.021 0.037 0.804 0.041 11.848 0.185 0.031 6.015 0.048 0.023 0.043 0.762 87 Figure 4.6: Tracking performance metrics across all configuration settings on the measured motion and MI data. evaluation measures of optimal models using the measured data for each setting. Representative samples of motion tracking in all settings are also displayed in Figure 4.7. Our results indicate that the parallel setting with the optimal regression model outperforms other settings on both measured and synthetic MI data. 4.5 Magnetic Induction vs. RFID 4.5.1 Background An RFID system consists of a reader and one or multiple tags, also known as transponders, com- monly used to automatically detect and track objects within its proximity range. It works based on the fundamental principles and concepts of electromagnetic fields and inductive coupling to trans- mit data and information. It has been implemented in a diverse range of areas, from item tracking to medical applications. RFID tags can be divided into two categories of passive and active. Active tags require a power source and are either integrated with a battery or a power infrastructure. In this case, a tag’s lifetime is limited by its source/battery, which makes active tags impractical for applications employing small and portable devices. On the other hand, passive tags don’t require batteries or maintenance, which provides them little with an unlimited operational lifetime, 88 (b) (a) (b) Figure 4.7: Distance and motion tracking in the z-direction. 89 In this scenario, the RFID reader is responsible for communication and powering the passive tags. The reader sends an interrogation signal to the title that energizes the transponder. The tag activates and sends back its unique identifier (UID) if the received power is higher than its sensitivity [141]. The power/data transfer between the reader and a remote passive tag can be done via far-field, and near-field-based signals [142]. The EM wave generated by the antenna divides into near-field (region close to the antenna within the range of a wavelength) and far-field. The radiative far-field has electric and magnetic components, while the non-radiative near-field is mainly magnetic. Far-field tags usually operate in the 860-960 MHz ultra-high frequency (UHF) band or the 2.45 GHz Microwave band, and near-field coupling tags in 125 kHz low frequency (LF) and 13.56 MHz high frequency (HF). Many solutions and standards are available for RFIDs working in different frequency bands, such as ISO 15693 and 14443, proposed for the near-field approach [142]. The tags using near-field coupling for communication with reader work based on Faraday’s principle, similar to magnetic induction explained before. A reader sends an alternating current through the coil, generating an alternating magnetic field around it. It induces an alternating voltage at the tag’s coil near the reader antenna that can power the tag chip. These tags send back data using load modulation. A modulation resistance connected in parallel with the tag antenna switches between two different (usually conjugate matching and a short circuit) load impedances at the clock rate of the signal transmitted from the reader to modulate the signal [143]. Then the reader retrieves this signal by monitoring the change of current flowing through its coil. The magnetic field decays proportional to 1=r 3 , where r is the distance between the tag and reader, and inductive coupling works in the range of a wavelength. As a result, the coverage range of near-field coupling decreases as the frequency of operation increases. On the other hand, some applications that require a higher data rate have to operate in higher frequencies. For these applications, passive RFID tags working based on far-field communication are a suitable solution. A far-field-based passive tag receives the propagating EM waves radiated from the reader antenna and powers itself. These types of tags send back data using the backscattering technique instead of 90 load modulation. The antenna can be tuned to a specific frequency to absorb the received energy at that frequency. Hence, a signal can be encoded by changing the antenna’s impedance over time [144]. The received signal strength indicator (RSSI) value that an RFID reader reports determines the power level of the signal returned from a tag and shows how well it responds. RSSI is usually measured in dBm or decibels per milliwatt. Several factors can affect the RSSI value, including system parameters (e.g., coil’s radius, number of turns, etc.) and environmental effects. One of these parameters is the distance between the tag and the reader, making it a distinctive feature for detecting tagged objects. Studies have proposed UHF RFID-based localization methods. For example, [145] surveys RFID-based localization techniques, their challenges, positioning princi- ples. [146] provides an introduction to the state-of-the-art of UHF RFID localization methods and applications. It categorizes the techniques into two major types: range-based and range-free, and presents a comparative study of surveyed works and trends. In far-field radiating RFID systems; however, the environmental parameters are another influ- ential parameter. It makes the motion tracking based on only the signal strength to be challenging. For example, RSSI value reported in an UHF system is subject to the background medium’s per- meability. Moreover, the reader radiates signal into space (even far-field), and the multipath effect caused by signals reflecting off surfaces affects the measurements. Most works obtain the distance or angle between antenna and tag utilizing the RSSI combined with other features such as time of arrival (ToA), phase of arrival (PoA), or angle of arrival (AoA). The work presented in [147], for example, points out the challenges of RSSI-based localization methods using passive tags, includ- ing high complexity and low accuracy. It presents a method based on AoA, estimated by phase difference, to locate tags accurately. The majority of available research works RFID positioning systems exploit far-field signals in UHF bands. However, the lower frequency systems exploiting magnetic induction coupling can remove environmental effects and provide more information for positioning, similar to the proposed MI system for motion tracking. Moreover, the objective of most studies is to locate 91 static objects rather than tracking in-motion targets. For moving object scenarios, one or multiple antennas are fixed at predefined locations and retrieve motion within the coverage range of readers [125]. Here, we investigate the capability of HF passive RFID system in motion tracking and compare it with the proposed MI system. We compare the performance of systems using the strength of the signal returned from tag to the reader since the measured MI signals are the only features used in the MI-based method. 4.5.2 Measurement Setup To compare RSSI and MI signals relationship with motion data, we performed the experiments explained in section 4.3 using HF RFID tags instead of MI sensors. Industrial passive tags are usually built as integrated circuits (ICs), which can be attached to a customized antenna. Here we adopted the ST25DV04K chip manufactured by STMicroelectronics, which is a dynamic near- field communication (NFC) RFID tag with 4 Kbit electrically erasable programmable memory (EEPROM) and fast transfer mode (FTM) capability. This dynamic tag uses ISO 15693 and ISO 14443 protocols based on passive RFID technology and operates in the HF range, at 13.56 MHz. The tag gets energized when placed in the reader’s magnetic field, enabling its built-in circuitry to demodulate the data transmitted from the reader. The RFID reader keeps the magnetic field at the end of requests without any modulation to power the tag, enabling it to send back its reply to the reader. The tag chip sends back its response by internally modulating its input impedance (load modulation) In order to design the circuit of a passive tag, its resonance frequency must match the reader operating frequency (13.56 MHz) since the tag obtains its power from the generated field by the reader. The power transfer from the reader to the tag and the communication range maximizes, as the resonant frequency tunes to the carrier frequency. Hence, the tag antenna must be designed such that its equivalent inductance matches the internal tuning capacitance value to build a circuit resonating at the resonance frequency. The tuning capacitance is typically specified in the user manual by the manufacturer. The tuning capacitance for the ST25DV04K chip is 28.5 pF. We 92 (a) (b) Figure 4.8: The designed RFID tag with customized antenna and variable capacitor. designed an air-cored, three-layer copper coil with a 5 cm radius and 34 AWG wire diameter as an antenna to be attached to the tag integrated circuit (IC). The wires are taped into a plastic ring to make fix the antenna inductance and unchanged over time. We also used a variable capacitor in parallel to the tag IC to adjust the equivalent capacitance to the inductance of the antenna. Figure 4.8 shows the RFID tag with its customized antenna and variable capacitor designed to operate at 13.56 MHz. Different parameters can impact the tuning frequency of the tag antenna. Hence we tested it after development by using a network analyzer and a loop probe. A self-made single turn loop similar to the tag antenna with a copper wire and a coaxial connector is used in the measurements. The loop probe is then connected to the output of the VNA set in reflection mode to measure S11. The probe and tag mutually couple when the antenna under the test is placed within the loop probe fields, which changes the loop probe impedance. At the resonant frequency of the tag, the S11 value reaches its minimum and can be used to tune the matching by changing the variable capacitor. 93 Figure 4.9: An experimental RFID measurement including RSSI and motion data. 4.5.3 Results We measured RSSI data reported from the reader using FEIG’s ISOStart software with an average sampling interval of 75 ms. The Kinect camera is also utilized to capture the motion of the reader and transponders while the reader records their RSSI values. The experiments are performed for 112 different movements with a duration of about 30 seconds for each sample. Both motion data recorded by Kinect (as explained in Section 4.3.1) and RSSI data are resampled with the time interval of 0.1 s. Figure 4.9 displays a sample including its measured RSSI and motions data. Using statistical measures, including correlation and R 2 , we studied the relationship between the RSSI data and the 3D distance of its corresponding tag antenna from the reader. Then the results are compared with their equivalent values calculated for MI data. These statistical measures are calculated for the whole dataset obtained from MI and RFID experiments. Figure 4.10 presents the average R 2 and the correlation for both experiments. As the results indicate, the MI signal has a stronger relationship with its motion compared to a passive tag. The calculated statistical express that movements in the Z-direction, which is the alignment of the reader antenna, have more effect 94 Figure 4.10: Comparison of MI signal and RSSI of passive RFID in 3D motion tracking using statistical measures. on the MI signal than other directions. The outcomes are consistent with the previously presented results in Section 4.4.2 showing the accuracy of motion reconstruction using MI data is higher in the direction of reader surface norm. The RSSI value measured by the reader is sensitive to various system parameters, including variation in the load at the tag side for load modulation. As explained before, the load at the tag side varies, typically between matching load and short circuit, to send back data [148]. In other words, the tag coil is short-circuited for a certain portion of the communication time. As the power delivery between the reader and tag is largely affected by the tag load, the average returned power no longer remains sensitive to the movement as before. Other parameters such as the number of ones and zeros, as a result, can affect the backscattering power from tag to the reader, making RSSI value a weak descriptive feature of an RFID tag movement. We evaluated several machine learning regression models on the RFID experimental measure- ments to track the distance of a passive tag using its recorded RSSI at the reader. Here we only trained models on the measured RSSI data instead of synthetic signals since the MI model is not able to estimate the returned power from a tag to the reader using their motion data. The reason is that the load at the tag side, which is an important input parameter of the MI model, varies over time (load modulation). Therefore, the amount of data available for model training and evaluation are restricted by the number of performed experiments. Due to the limited number of measurements, complex machine learning models are not implemented. The models are implemented and com- 95 Table 4.2: Performance of regression models in motion tracking using RSSI data measured by the RFID reader. Model RMSE MAPE R2 LightGBM 0.062 0.137 0.284 RF 0.062 0.134 0.273 ET 0.063 0.134 0.26 KNN 0.063 0.135 0.249 MLP 0.064 0.145 0.225 LR 0.067 0.153 0.151 DT 0.085 0.169 0.102 pared using the PyCaret library, and the performance results are presented in Table 4.2. Distance between tag and coils are calculated using their measured RSSI over time while the tag moves. The presented performance results indicate that the signal strength returned to the reader are not proper features for tracking passive tags. It should also be noted that the models are trained on smaller datasets (measured data) compared to the number of the synthetic data used for training MI models, which can also degrade the performance of regression models. 4.6 Discussion We demonstrated an MI-based system to precisely track the motion of receivers with respect to the central transmitter node. We employed an HF-RFID transmitter module equipped with a loop antenna. A simple integrated circuit is also used as a receiver, which interfaces with the central node, and records received signals sent from the transmitter using an Arduino. The main focus of this chapter was to provide a proof of concept for the proposed system, which can also be implemented for real-world applications by proper modifications. For example, the MI coils should be designed to be suitable for wearing on the human wrist, arm, and ankle. Furthermore, a wearable custom-designed central node capable of driving a controlled amount of current at the operating frequency through its coil is required. The receivers should cover the range of about 0.5 m to 1 m with minimum power consumption. The RF output power of the reader we used here is 1 Watt, which can be reduced by designing an efficient capable of sensing lower power signals. 96 For example, [149] presents a transceiver design exploiting the low path loss of Magnetic Human Body Communication (mHBC) communication channels toward ultra-efficient body area networking. The transmitter and receiver respectively require only 7.15 and 4.7 pJ/bit, and their design is a helpful reference for implementing MI transceivers. One implementation approach to reduce the number of nodes with battery is to make the central node serve as both transmitter and receiver. It means that the central unit can broadcast the signal and listens back to the responses reflected from the sensors. It is similar to the working principle of an RFID system based on pas- sive (battery-less) tags explained in Section 4.5.1. Using the RFID protocol, the central node can communicate with the sensors via a secure near-field link backscattering from them. Although the receiver nodes can be realized as passive reflectors similar to passive RFID tags, load modulation is not a practical solution for data transmission in an MI system. As explained before, the backscattered field, and consequently, the voltage signal received by the reader, switches over two values[121]. The average power returned to the reader is no longer a direct function of distance and misalignment between coils since it varies by the number of zeros and ones in the data stream. The performance analysis and comparison results presented in this chapter point out the important parameters that should be considered toward designing an MI system-based motion tracking system. The system relies on the variation of transferred power between transmitter and sensors. Hence, the parameters that can significantly affect the sensitivity of the system must be eliminated. The principle underlying many existing technologies such as RFIDs and power transfer systems is similar to the proposed MI system. The technological developments available for these systems can help design a practical and ready-to-use MI system with proper modification and consideration. 4.7 Conclusion We presented an MI sensor designed for 3D motion tracking representing the movements by vari- ation in the signals received from the transmitter unit instead of measuring spatial data. This approach can overcome obstacles associated with conventional wearable sensor-based systems, 97 such as reducing power consumption, minimizing the effects of environmental interference, and achieving higher tracking accuracy. We employed machine learning regression models and investi- gated their performance using synthetic and measured data. The models are trained on synthesized data generated by an analytical model, which removes the need for supervised training data. The synthesizing model must be calibrated only once to scale the synthetic training data to sensor mea- surements and tune the regression algorithm. The system with an optimal regressor can track with a mean error of 3 cm, evaluated on real-world measured data. 98 Chapter 5 Conclusion and Future Work We introduced MI-based communication technique for WBANs and studied its performance around the human. An analytical model is then proposed for easier system design and power delivery anal- ysis by adopting the circuit model of small loop antennas and mutual inductance between two coils. The mutual inductance between coils is a key parameter in computing the gain of MI communi- cation system and power transmission analysis. Therefore, we derived an expression for mutual inductance between two arbitrarily distanced and oriented coils, considering frequency, coil ge- ometry, misalignment, distance, and environment. It means that the MI model can be applied for different scenarios, in which the distance and misalignment of the transmitter and receiver are taken into account. The proposed model is validated using HFSS simulations and also experimental measurements. The experiments are performed for coils with different structural parameters, including radius, wire AWG, and the number of turns. To study the effect of the biological tissues on system performance, the measurements and simulations are conducted with and without the presence of a human body. We compared the forward voltage gain (S 21 ) calculated by the HFSS simulations and the measured values obtained from VNA with their corresponding results estimated by the proposed MI model. We showed that the proposed analytical model is valid for the frequency lower than 30 MHz with 99 much faster speed in gain calculation than full-wave simulators such as HFSS. Compared to the wavelength of higher frequencies, the dimension of the human body is not negligible anymore and can affect the system performance. A general framework for human activity detection and tracking motion in 3D using MI system is then presented. This method can tackle challenges of conventional HAR systems in terms of, for example, power consumption and cost. The MI-based motion tracking system also works well in non-magnetic harsh environments. It is a suitable system for applications operating in these situations, such as monitoring scuba-diver underwater. We provided a proof-of-concept for the proposed system using synthetic data, measurements, and machine learning algorithms. The applicability of the system in detecting human movements is evaluated by simulating the MI data of the system assumed to be attached to a user performing different activities. The MI signals of eight transmitters communicating with a central node attached to the torso are generated using MI model and two publicly available MoCap datasets. The synthesized data are used as inputs for several commonly used machine learning classifiers and deep recurrent neural networks to detect user’s activities. We presented a deep recurrent neural network framework for human activity detection using MI signals, which outperforms other classifiers considered in this work. We also compared the accuracy of the proposed MI-HAR system with other state-of-the-art methods using different modalities for activity detection. In addition to activity detection, the system is evaluated for tracking motion in 3D space. We developed a prototype sensor for experimental measurement and assessed the performance of three different sensor configurations. In each experiment, the MI signals received at the moving node are recorded using Arduino, and the 3D motion of the node is captured using a Kinect camera as the ground-truth motion data. Using machine learning regression algorithms and recorded MI data, we aimed to reconstruct the node movements. We adopted V AE and the MI model to generate more data for training the regression models and enhance the performance. It removes the need for labeled measured data containing MI and 3D motion data. The performance of regression 100 models is investigated on the synthetic data, and the optimal models are then tested on the real- world samples. The outcomes demonstrate that the proposed system is able to accurately track the movements of an MI node. Other low-power sensors, such as inertial units, are also available that can be combined with the proposed MI sensor to enhance the MI system accuracy in human activity recognition and motion tracking. By including biodata and environmental features, the system can be used in many more monitoring applications. The extension of this work is designing a wearable sensor network, including hardware and software platforms for fusing different sensors and collecting their data in real-time on a central node. In order to take a step towards commercialization, wearable devices and system nodes have to be miniaturized by designing low-power integrated circuits. Devices can be optimized for a specific application, such as tracking disease states and progression. The designed system must have the ability to interface with the sensors and communicating amongst distributed nodes via a secure wireless link based on magnetic induction to obtain sensor data. The protocols and standards available for NFC/RFID system are useful sources to design different layer of the on-body communication system. However, previously mentioned factors affecting the sensitivity of the MI system must be taken into account for developing a suitable communication technique. The collected data at the central node from sensors can either be processed on or off the body. In the on-body case, the central unit process and fuses sensor data in addition to the data collection task. In the latter case, the collected data are transferred wirelessly to an off-body device, e.g., a smartphone, for processing and decision making. The off-body communication link must cover a longer range than the MI-based method. Low-power wireless technologies such as BLE are suitable options. Another aspect of this work that needs further research is data analysis methods. Co-design of sensor platform and processing algorithms can significantly maximize efficiency, performance, and inference accuracy. Hence, developing and evaluating machine learning and neural network models for a targeted application is another stage of system design procedure that one must investigate. Since the analysis and processing algorithms are executed on-device, the 101 algorithms must be optimized for on-chip execution. Future work is to develop and evaluate basic algorithms for data cleaning, feature extraction, fusion, etc., considering the available power and memory of portable devices to achieve the desired accuracy. 102 Bibliography [1] H. Cao, V . Leung, C. Chow, and H. Chan, “Enabling technologies for wireless body area networks: A survey and outlook,” IEEE Communications Magazine, vol. 47, no. 12, pp. 84–93, 2009. [2] S. Ullah, P. Khan, N. Ullah, S. Saleem, H. Higgins, and K. S. Kwak, “A review of wireless body area networks for medical applications,” arXiv preprint arXiv:1001.0831, 2010. [3] M. Patel and J. Wang, “Applications, challenges, and prospective in emerging body area networking technologies,” IEEE Wireless Communications Magazine, vol. 17, no. 1, pp. 80–88, 2010. [4] E. Jovanov, A. Milenkovic, C. Otto, and P. C. De Groen, “A wireless body area network of intelligent motion sensors for computer assisted physical rehabilitation,” Journal of Neuro- Engineering and Rehabilitation, vol. 2, no. 1, p. 6, 2005. [5] M. Karunanithi, “Monitoring technology for the elderly patient,” Expert review of medical devices, vol. 4, no. 2, pp. 267–277, 2007. [6] S. Kumar, K. Kambhatla, F. Hu, M. Lifson, and Y . Xiao, “Ubiquitous computing for remote cardiac patient monitoring: a survey,” International journal of telemedicine and applica- tions, vol. 2008, p. 3, 2008. [7] National institutes of health. [Online]. Available: https://www.nih.gov/ [8] J. M. Ortman, V . A. Velkoff, H. Hogan et al., “An aging nation: the older population in the united states,” Washington, DC: US Census Bureau, pp. 25–1140, 2014. [9] Y . Hao and R. Foster, “Wireless body sensor networks for health-monitoring applications,” Physiological measurement, vol. 29, no. 11, p. R27, 2008. [10] A. Darwish and A. E. Hassanien, “Wearable and implantable wireless sensor network solu- tions for healthcare monitoring,” Sensors, vol. 11, no. 6, pp. 5561–5595, 2011. [11] World health organization. [Online]. Available: http://www.who.int/en/ [12] A. Milenkovi´ c, C. Otto, and E. Jovanov, “Wireless sensor networks for personal health monitoring: Issues and an implementation,” Computer communications, vol. 29, no. 13, pp. 2521–2533, 2006. 103 [13] Y . Lu, Y . Wei, L. Liu, J. Zhong, L. Sun, and Y . Liu, “Towards unsupervised physical activity recognition using smartphone accelerometers,” Multimedia Tools and Applications, vol. 76, no. 8, pp. 10 701–10 719, 2017. [14] B. Cvetkovi´ c, R. Szeklicki, V . Janko, P. Lutomski, and M. Luˇ strek, “Real-time activity monitoring with a wristband and a smartphone,” Information Fusion, vol. 43, pp. 77–93, 2018. [15] M. M. Hassan, M. Z. Uddin, A. Mohamed, and A. Almogren, “A robust human activity recognition system using smartphone sensors and deep learning,” Future Gener. Comp. Sy., vol. 81, pp. 307–313, (FUTURE GENER COMP SY 2018. [16] J. Wang, Y . Chen, S. Hao, X. Peng, and L. Hu, “Deep learning for sensor-based activity recognition: A survey,” Pattern Recognition Letters, vol. 119, pp. 3–11, 2019. [17] G. M. Weiss, J. L. Timko, C. M. Gallagher, K. Yoneda, and A. J. Schreiber, “Smartwatch- based activity recognition: A machine learning approach,” in 2016 IEEE-EMBS Interna- tional Conference on Biomedical and Health Informatics (BHI). IEEE, 2016, pp. 426–429. [18] S. Ha, S. Park, H. Lim, S. H. Baek, D. K. Kim, and S.-H. Yoon, “The placement position optimization of a biosensor array for wearable healthcare systems,” Journal of Mechanical Science and Technology, vol. 33, no. 7, pp. 3237–3244, 2019. [19] H. Allahbakhshi, T. Hinrichs, H. Huang, and R. Weibel, “The key factors in physical activity type detection using real-life data: A systematic review,” Front. Physiol., vol. 10, p. 75, 2019. [20] L. Chen, J. Hoey, C. D. Nugent, D. J. Cook, and Z. Yu, “Sensor-based activity recognition,” IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), vol. 42, no. 6, pp. 790–808, 2012. [21] M. Cornacchia, K. Ozcan, Y . Zheng, and S. Velipasalar, “A survey on activity detection and classification using wearable sensors,” IEEE Sensors Journal, vol. 17, no. 2, pp. 386–403, 2016. [22] L. Gao, A. Bourke, and J. Nelson, “Evaluation of accelerometer based multi-sensor ver- sus single-sensor activity recognition systems,” Medical Engineering and Physics, vol. 36, no. 6, pp. 779–785, 2014. [23] T. Sztyler, H. Stuckenschmidt, and W. Petrich, “Position-aware activity recognition with wearable devices,” Pervasive and Mobile Computing, vol. 38, pp. 281–295, 2017. [24] ¨ O. F. Ertuˇ grul and Y . Kaya, “Determining the optimal number of body-worn sensors for human activity recognition,” Soft Computing, vol. 21, no. 17, pp. 5053–5060, 2017. [25] S. Sharma, M. Tripathi, and V . Mishra, “Survey paper on sensors for body area network in health care,” in International Conference on Emerging Trends in Computing and Communi- cation Technologies (ICETCCT). IEEE, 2017, pp. 1–6. 104 [26] A. Mosenia, S. Sur-Kolay, A. Raghunathan, and N. K. Jha, “Wearable medical sensor-based system design: A survey,” IEEE Transactions on Multi-Scale Computing Systems, vol. 3, no. 2, pp. 124–138, 2017. [27] R. Gravina, P. Alinia, H. Ghasemzadeh, and G. Fortino, “Multi-sensor fusion in body sensor networks: State-of-the-art and research challenges,” Information Fusion, vol. 35, pp. 68–80, 2017. [28] S. Movassaghi, M. Abolhasan, J. Lipman, D. Smith, and A. Jamalipour, “Wireless body area networks: A survey,” IEEE Commun. Surv. Tutor., vol. 16, no. 3, pp. 1658–1686, 2014. [29] O. D. Lara and M. A. Labrador, “A survey on human activity recognition using wearable sensors.” IEEE Commun. Surv. Tutor., vol. 15, no. 3, pp. 1192–1209, 2012. [30] J. Kim, A. S. Campbell, B. E.-F. de ´ Avila, and J. Wang, “Wearable biosensors for healthcare monitoring,” Nature Biotechnology, vol. 37, no. 4, pp. 389–406, 2019. [31] G. Aroganam, N. Manivannan, and D. Harrison, “Review on wearable technology sensors used in consumer sport applications,” Sensors, vol. 19, no. 9, p. 1983, 2019. [32] R. Cavallari, F. Martelli, R. Rosini, C. Buratti, and R. Verdone, “A survey on wireless body area networks: Technologies and design challenges,” IEEE Commun. Surv. Tutor., vol. 16, no. 3, pp. 1635–1657, 2014. [33] R. Negra, I. Jemili, and A. Belghith, “Wireless body area networks: Applications and tech- nologies,” Procedia Computer Science, vol. 83, pp. 1274–1281, 2016. [34] N. Golestani and M. Moghaddam, “Theoretical modeling and analysis of magnetic induc- tion communication in wireless body area networks (wban),” IEEE Journal of Electromag- netics, RF and Microwaves in Medicine and Biology, vol. 2, no. 1, pp. 48–55, 2018. [35] K. Hasan, K. Biswas, K. Ahmed, N. S. Nafi, and M. S. Islam, “A comprehensive review of wireless body area network,” Journal of Network and Computer Applications, vol. 143, pp. 178–198, 2019. [36] D.-T. Huynh and M. Chen, “An energy efficiency solution for wban in healthcare monitoring system,” in 2016 3rd International Conference on Systems and Informatics (ICSAI). IEEE, 2016, pp. 685–690. [37] D. M. Barakah and M. Ammad-uddin, “A survey of challenges and applications of wire- less body area network (wban) and role of a virtual doctor server in existing architecture,” in 2012 Third International Conference on Intelligent Systems Modelling and Simulation. IEEE, 2012, pp. 214–219. [38] J. Bae, H. Cho, K. Song, H. Lee, and H.-J. Yoo, “The signal transmission mechanism on the surface of human body for body channel communication,” IEEE Transactions on Microwave Theory and Techniques, vol. 60, no. 3, pp. 582–593, 2012. [39] D. H. Werner and Z. H. Jiang, Electromagnetics of body area networks: antennas, propaga- tion, and RF systems. John Wiley and Sons, 2016. 105 [40] T. Hayajneh, G. Almashaqbeh, S. Ullah, and A. V . Vasilakos, “A survey of wireless tech- nologies coexistence in wban: analysis and open research issues,” Wireless Networks, vol. 20, no. 8, pp. 2165–2199, 2014. [41] R. Natarajan, P. Zand, and M. Nabi, “Analysis of coexistence between ieee 802.15. 4, ble and ieee 802.11 in the 2.4 ghz ism band,” in IECON 2016-42nd Annual Conference of the IEEE Industrial Electronics Society. IEEE, 2016, pp. 6025–6032. [42] N. Golestani and M. Moghaddam, “Human activity recognition using magnetic induction- based motion signals and deep recurrent neural networks,” Nature Communications, vol. 11, no. 1, pp. 1–11, 2020. [43] A. K. Sharma, S. Yadav, S. N. Dandu, V . Kumar, J. Sengupta, S. B. Dhok, and S. Kumar, “Magnetic induction-based non-conventional media communications: A review,” IEEE Sen- sors Journal, vol. 17, no. 4, pp. 926–940, 2016. [44] M. A. Hanson, H. C. Powell Jr, A. T. Barth, K. Ringgenberg, B. H. Calhoun, J. H. Aylor, and J. Lach, “Body area sensor networks: Challenges and opportunities,” Computer, vol. 42, no. 1, p. 58, 2009. [45] J. I. Agbinya, Principles of Inductive Near Field Communications for Internet of Things. River Publishers, 2011, vol. 18. [46] N. Golestani and M. Moghaddam, “Magnetic induction communications for wireless body area network,” in The National Radio Science Meeting, Boulder, Colorado, January 2017. [47] J. Heide, T. D. Gooch, A. D. Prescott, T. W. Sander, J. Davidson, and E. A. Renz, “Magnetic induction hearing aid,” Google Patents, January 1989. [48] J. I. Agbinya and M. Masihpour, “Power equations and capacity performance of mag- netic induction body area network nodes,” in Broadband and Biomedical Communications (IB2Com), 2010 Fifth International Conference on. IEEE, 2010, pp. 1–6. [49] J. I. Agbinya, N. Selvaraj, A. Ollett, S. Ibos, Y . Ooi-Sanchez, M. Brennan, and Z. Chaczko, “Size and characteristics of the ‘cone of silence’in near-field magnetic induction communi- cations,” power, vol. 1, p. 6, 2010. [50] S. Hoskins, T. Sobering, D. Andresen, and S. Warren, “Near-field wireless magnetic link for an ingestible cattle health monitoring pill,” in Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International Conference of the IEEE. IEEE, 2009, pp. 5401–5404. [51] S. Movassaghi, P. Arab, and M. Abolhasan, “Wireless technologies for body area networks: Characteristics and challenges,” in Communications and Information Technologies (ISCIT), 2012 International Symposium on. IEEE, 2012, pp. 42–47. [52] D. P. Kumar, T. Amgoth, and C. S. R. Annavarapu, “Machine learning algorithms for wire- less sensor networks: A survey,” Information Fusion, vol. 49, pp. 1–25, 2019. 106 [53] M. A. Alsheikh, S. Lin, D. Niyato, and H.-P. Tan, “Machine learning in wireless sensor networks: Algorithms, strategies, and applications,” IEEE Communications Surveys & Tu- torials, vol. 16, no. 4, pp. 1996–2018, 2014. [54] H. F. Nweke, Y . W. Teh, M. A. Al-Garadi, and U. R. Alo, “Deep learning algorithms for human activity recognition using mobile and wearable sensor networks: State of the art and research challenges,” Expert Syst. Appl., vol. 105, pp. 233–261, 2018. [55] N. Golestani and M. Moghaddam, “Communication system design for magnetic induction- based wireless body area network,” in USNC-URSI Radio Science Meeting (Joint with AP-S Symposium). IEEE, 2017, pp. 49–50. [56] N. Tal, Y . Morag, L. Shatz, and Y . Levron, “Design optimization of transmitting antennas for weakly coupled magnetic induction communication systems,” PloS one, vol. 12, no. 2, p. e0171982, 2017. [57] R. Bansal, “Near-field magnetic communication,” IEEE Antennas Propag. Mag., vol. 46, no. 2, pp. 114–115, 2004. [58] Z. Sun and I. F. Akyildiz, “Magnetic induction communications for wireless underground sensor networks,” IEEE Transactions on Antennas and Propagation, vol. 58, no. 7, pp. 2426–2435, 2010. [59] C. Bunszel, “Magnetic induction: A low-power wireless alternative,” RF DESIGN, vol. 24, no. 11, pp. 78–80, 2001. [60] J. I. Agbinya and M. Masihpour, “Power equations and capacity performance of magnetic induction communication systems,” Wireless Personal Communications, vol. 64, no. 4, pp. 831–845, 2012. [61] Z. Sun, I. F. Akyildiz, S. Kisseleff, and W. Gerstacker, “Increasing the capacity of magnetic induction communications in rf-challenged environments,” IEEE Transactions on Commu- nications, vol. 61, no. 9, pp. 3943–3952, 2013. [62] H.-J. Kim, H. Hirayama, S. Kim, K. J. Han, R. Zhang, and J.-W. Choi, “Review of near-field wireless power and communication for biomedical applications,” IEEE Access, vol. 5, pp. 21 264–21 285, 2017. [63] M. C. Domingo, “Magnetic induction for underwater wireless communication networks,” IEEE Transactions on Antennas and Propagation, vol. 60, no. 6, pp. 2929–2939, 2012. [64] C. A. Balanis, Antenna theory: analysis and design. John Wiley & Sons, 2016. [65] A. R. Silva and M. Moghaddam, “Operating frequency selection for low-power magnetic induction-based wireless underground sensor networks,” in 2015 IEEE Sensors Applications Symposium (SAS). IEEE, 2015, pp. 1–6. [66] D. K. Cheng et al., Field and Wave Electromagnetics, ser. Pearson Education India. Addison-Wesley, 1989. 107 [67] D. Cheng, Field and Wave Electromagnetics, ser. Addison-Wesley series in electrical engi- neering. Addison-Wesley, 1983. [68] A. R. Silva and M. Moghaddam, “Design and implementation of low-power and mid-range magnetic-induction-based wireless underground sensor networks,” IEEE Transactions on Instrumentation and Measurement, vol. 65, no. 4, pp. 821–835, 2016. [69] E. B. Rosa, Calculation of the self-inductance of single-layer coils. US Government Print- ing Office, 1906, no. 31. [70] N. Golestani and M. Moghaddam, “Improving the efficiency of magnetic induction-based wireless body area network (wban),” in 2018 IEEE International Microwave Biomedical Conference (IMBioC). IEEE, 2018, pp. 166–168. [71] S. J. Orfanidis, Electromagnetic waves and antennas. Rutgers University New Brunswick, NJ, 2002. [72] D. A. Frickey, “Conversions between s, z, y, h, abcd, and t parameters which are valid for complex source and load impedances,” IEEE Transactions on Microwave Theory and Techniques, vol. 42, no. 2, pp. 205–211, 1994. [73] J. Wang and Q. Wang, Body area communications: channel modeling, communication sys- tems, and EMC. John Wiley & Sons, 2012. [74] S. Gabriel, R. Lau, and C. Gabriel, “The dielectric properties of biological tissues: 3- parametric models for the dielectric spectrum of tissues,” Physics in medicine and biology, vol. 41, no. 11, p. 2271, 1996. [75] ——, “The dielectric properties of biological tissues: 2. measurements in the frequency range 10 hz to 20 ghz,” Physics in medicine and biology, vol. 41, no. 11, p. 2251, 1996. [76] J. Yin, Q. Yang, and J. J. Pan, “Sensor-based abnormal human-activity detection,” IEEE Transactions on Knowledge and Data Engineering, no. 8, pp. 1082–1090, 2008. [77] J. Kim, M. Kim, M.-S. Lee, K. Kim, S. Ji, Y .-T. Kim, J. Park, K. Na, K.-H. Bae, H. K. Kim et al., “Wearable smart sensor systems integrated on soft contact lenses for wireless ocular diagnostics,” Nature Communications, vol. 8, no. 1, pp. 1–8, 2017. [78] S. C. Mukhopadhyay, “Wearable sensors for human activity monitoring: A review,” IEEE Sensors Journal, vol. 15, no. 3, pp. 1321–1330, 2014. [79] S. E. Jones, V . T. van Hees, D. R. Mazzotti, P. Marques-Vidal, S. Sabia, A. van der Spek, H. S. Dashti, J. Engmann, D. Kocevska, J. Tyrrell et al., “Genetic studies of accelerometer- based sleep measures yield new insights into human sleep behaviour,” Nature Communica- tions, vol. 10, no. 1, pp. 1–12, 2019. [80] W. Wang, A. X. Liu, M. Shahzad, K. Ling, and S. Lu, “Understanding and modeling of wifi signal based human activity recognition,” in Proceedings of the 21st annual international conference on mobile computing and networking, 2015, pp. 65–76. 108 [81] R. Poppe, “A survey on vision-based human action recognition,” Image and Vision Comput- ing, vol. 28, no. 6, pp. 976–990, 2010. [82] H.-B. Zhang, Y .-X. Zhang, B. Zhong, Q. Lei, L. Yang, J.-X. Du, and D.-S. Chen, “A com- prehensive survey of vision-based human action recognition methods,” Sensors, vol. 19, no. 5, p. 1005, 2019. [83] R. Bodor, B. Jackson, and N. Papanikolopoulos, “Vision-based human tracking and activity recognition,” in 11th Mediterranean Conference on Control and Automation, vol. 1, 2003. [84] T. B. Moeslund, A. Hilton, and V . Kr¨ uger, “A survey of advances in vision-based human motion capture and analysis,” Comput. Vis. Image Und., vol. 104, no. 2-3, pp. 90–126, 2006. [85] P. Kumari, L. Mathew, and P. Syal, “Increasing trend of wearables and multimodal interface for human activity monitoring: A review,” Biosens. Bioelectron., vol. 90, pp. 298–307, 2017. [86] A. Hua, Z. Quicksall, C. Di, R. Motl, A. Z. LaCroix, B. Schatz, and D. M. Buchner, “Accelerometer-based predictive models of fall risk in older women: a pilot study,” NPJ Digital Medicine, vol. 1, no. 1, pp. 1–8, 2018. [87] “Vicon,” https://www.vicon.com/hardware/blue-trident, accessed: 2020-01-29. [88] “Zenshin technology,” https://zenshin-tech.com, accessed: 2020-01-29. [89] I. Prayudi and D. Kim, “Design and implementation of imu-based human arm motion cap- ture system,” in IEEE International Conference on Mechatronics and Automation. IEEE, 2012, pp. 670–675. [90] R. Vemulapalli, F. Arrate, and R. Chellappa, “Human action recognition by representing 3d skeletons as points in a lie group,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 588–595. [91] Y . Ma, H. M. Paterson, and F. E. Pollick, “A motion capture library for the study of identity, gender, and emotion perception from biological motion,” Behav. Res. Methods, vol. 38, no. 1, pp. 134–141, 2006. [92] F. Ofli, R. Chaudhry, G. Kurillo, R. Vidal, and R. Bajcsy, “Berkeley mhad: A comprehensive multimodal human action database,” in IEEE Workshop on Applications of Computer Vision (WACV). IEEE, 2013, pp. 53–60. [93] R. Johari, J. V . Krogmeier, and D. J. Love, “Analysis and practical considerations in imple- menting multiple transmitters for wireless power transfer via coupled magnetic resonance,” IEEE Transactions on Industrial Electronic, vol. 61, no. 4, pp. 1774–1783, 2013. [94] B. L. Cannon, J. F. Hoburg, D. D. Stancil, and S. C. Goldstein, “Magnetic resonant cou- pling as a potential means for wireless power transfer to multiple small receivers,” IEEE Transactions on Power Electronics, vol. 24, no. 7, pp. 1819–1825, 2009. 109 [95] A. Mimouna, A. B. Khalifa, and N. E. B. Amara, “Human action recognition using triaxial accelerometer data: selective approach,” in 15th International Multi-Conference on Systems, Signals & Devices (SSD). IEEE, 2018, pp. 491–496. [96] F. J. Ord´ o˜ nez and D. Roggen, “Deep convolutional and lstm recurrent neural networks for multimodal wearable activity recognition,” Sensors, vol. 16, no. 1, p. 115, 2016. [97] N. Y . Hammerla, S. Halloran, and T. Ploetz, “Deep, convolutional, and recurrent models for human activity recognition using wearables,” arXiv preprint arXiv:1604.08880, 2016. [98] C. M. Bishop, Pattern recognition and machine learning. springer, 2006. [99] K. P. Murphy, Machine learning: a probabilistic perspective. MIT press, 2012. [100] F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg et al., “Scikit-learn: Machine learning in python,” Journal of Machine Learning Research, vol. 12, no. Oct, pp. 2825–2830, 2011. [101] P.-N. Tan, M. Steinbach, and V . Kumar, “Classification: basic concepts, decision trees, and model evaluation,” Introduction to data mining, vol. 1, pp. 145–205, 2006. [102] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back- propagating errors,” nature, vol. 323, no. 6088, pp. 533–536, 1986. [103] I. Goodfellow, Y . Bengio, A. Courville, and Y . Bengio, Deep learning. MIT press Cam- bridge, 2016, vol. 1. [104] W. Zhu, C. Lan, J. Xing, W. Zeng, Y . Li, L. Shen, and X. Xie, “Co-occurrence feature learning for skeleton based action recognition using regularized deep lstm networks,” arXiv preprint arXiv:1603.07772, 2016. [105] J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” arXiv preprint arXiv:1412.3555, 2014. [106] A. Murad and J.-Y . Pyun, “Deep recurrent neural networks for human activity recognition,” Sensors, vol. 17, no. 11, p. 2556, 2017. [107] H. Jaeger, Tutorial on training recurrent neural networks, covering BPPT, RTRL, EKF and the” echo state network” approach. GMD-Forschungszentrum Informationstechnik Bonn, 2002, vol. 5. [108] R. Pascanu, C. Gulcehre, K. Cho, and Y . Bengio, “How to construct deep recurrent neural networks,” arXiv preprint arXiv:1312.6026, 2013. [109] T.-Y . Pan, C.-H. Kuo, H.-T. Liu, and M.-C. Hu, “Handwriting trajectory reconstruction using low-cost imu,” IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 3, no. 3, pp. 261–270, 2018. 110 [110] D. Giansanti, G. Maccioni, F. Benvenuti, and V . Macellari, “Inertial measurement units furnish accurate trunk trajectory reconstruction of the sit-to-stand manoeuvre in healthy subjects,” Medical and Biological Engineering and Computing, vol. 45, no. 10, pp. 969– 976, 2007. [111] S. Suvorova, T. Vaithianathan, and T. Caelli, “Action trajectory reconstruction from inertial sensor measurements,” in 11th International Conference on Information Science, Signal Processing and their Applications (ISSPA). IEEE, 2012, pp. 989–994. [112] T. J. Shields, M. R. Amer, M. Ehrlich, and A. Tamrakar, “Action-affect-gender classifica- tion using multi-task representation learning,” in IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, 2017, pp. 2249–2258. [113] E. Garcia-Ceja, C. E. Galv´ an-Tejada, and R. Brena, “Multi-view stacking for activity recog- nition with sound and accelerometer data,” Information Fusion, vol. 40, pp. 45–56, 2018. [114] F. M. Noori, B. Wallace, M. Z. Uddin, and J. Torresen, “A robust human activity recog- nition approach using openpose, motion features, and deep recurrent neural network,” in Scandinavian Conference on Image Analysis. Springer, 2019, pp. 299–310. [115] F. S. Parizi, E. Whitmire, and S. Patel, “Auraring: Precise electromagnetic finger tracking,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 3, no. 4, pp. 1–28, 2019. [116] V . Camomilla, E. Bergamini, S. Fantozzi, and G. Vannozzi, “Trends supporting the in-field use of wearable inertial sensors for sport performance evaluation: A systematic review,” Sensors, vol. 18, no. 3, p. 873, 2018. [117] M. Ghislieri, L. Gastaldi, S. Pastorelli, S. Tadano, and V . Agostini, “Wearable inertial sen- sors to assess standing balance: A systematic review,” Sensors, vol. 19, no. 19, p. 4075, 2019. [118] C. Prakash, R. Kumar, and N. Mittal, “Recent developments in human gait research: pa- rameters, approaches, applications, machine learning techniques, datasets and challenges,” Artificial Intelligence Review, vol. 49, no. 1, pp. 1–40, 2018. [119] S. Ramasamy Ramamurthy and N. Roy, “Recent trends in machine learning for human ac- tivity recognition—a survey,” Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 8, no. 4, p. e1254, 2018. [120] A. Bux, P. Angelov, and Z. Habib, “Vision based human activity recognition: a review,” in Advances in Computational Intelligence Systems. Springer, 2017, pp. 341–371. [121] S. Zhang, Z. Wei, J. Nie, L. Huang, S. Wang, and Z. Li, “A review on human activity recognition using vision-based method,” Journal of healthcare engineering, vol. 2017, 2017. [122] Q. Lin, L. Yang, Y . Sun, T. Liu, X.-Y . Li, and Y . Liu, “Beyond one-dollar mouse: A battery- free device for 3d human-computer interaction via rfid tags,” in 2015 IEEE Conference on Computer Communications (INFOCOM). IEEE, 2015, pp. 1661–1669. 111 [123] I. Weygers, M. Kok, M. Konings, H. Hallez, H. De Vroey, and K. Claeys, “Inertial sensor- based lower limb joint kinematics: A methodological systematic review,” Sensors, vol. 20, no. 3, p. 673, 2020. [124] M. Zhao, Y . Tian, H. Zhao, M. A. Alsheikh, T. Li, R. Hristov, Z. Kabelac, D. Katabi, and A. Torralba, “Rf-based 3d skeletons,” in Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication, 2018, pp. 267–281. [125] C. Wang, J. Liu, Y . Chen, L. Xie, H. B. Liu, and S. Lu, “Rf-kinect: A wearable rfid-based approach towards 3d body movement tracking,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 2, no. 1, pp. 1–28, 2018. [126] A. Filippeschi, N. Schmitz, M. Miezal, G. Bleser, E. Ruffaldi, and D. Stricker, “Survey of motion tracking methods based on inertial sensors: A focus on upper limb human motion,” Sensors, vol. 17, no. 6, p. 1257, 2017. [127] M. Airaksinen, O. R¨ as¨ anen, E. Il´ en, T. H¨ ayrinen, A. Kivi, V . Marchi, A. Gallen, S. Blom, A. Varhe, N. Kaartinen et al., “Automatic posture and movement tracking of infants with wearable movement sensors,” Scientific reports, vol. 10, no. 1, pp. 1–13, 2020. [128] M. Al-Amri, K. Nicholas, K. Button, V . Sparkes, L. Sheeran, and J. L. Davies, “Inertial mea- surement units for clinical movement analysis: reliability and concurrent validity,” Sensors, vol. 18, no. 3, p. 719, 2018. [129] F. M. Pfister, T. T. Um, D. C. Pichler, J. Goschenhofer, K. Abedinpour, M. Lang, S. Endo, A. O. Ceballos-Baumann, S. Hirche, B. Bischl et al., “High-resolution motor state detec- tion in parkinson’s disease using convolutional neural networks,” Scientific reports, vol. 10, no. 1, pp. 1–11, 2020. [130] E. Garcia-Ceja, M. Riegler, T. Nordgreen, P. Jakobsen, K. J. Oedegaard, and J. Tørresen, “Mental health monitoring with multimodal sensing and machine learning: A survey,” Per- vasive and Mobile Computing, vol. 51, pp. 1–26, 2018. [131] H. Tannous, D. Istrate, A. Benlarbi-Delai, J. Sarrazin, D. Gamet, M. C. Ho Ba Tho, and T. T. Dao, “A new multi-sensor fusion scheme to improve the accuracy of knee flexion kinematics for functional rehabilitation movements,” Sensors, vol. 16, no. 11, p. 1914, 2016. [132] “Feig company,” https://www.feig-electronics.com/, accessed: 2020-09-20. [133] E. Lachat, H. Macher, T. Landes, and P. Grussenmeyer, “Assessment and calibration of a rgb-d camera (kinect v2 sensor) towards a potential use for close-range 3d modeling,” Remote Sensing, vol. 7, no. 10, pp. 13 070–13 097, 2015. [134] D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013. [135] N. Laptev, “Anogen: Deep anomaly generator,” in Outlier Detection De-constructed (ODD) Workshop. Available online: https://research. fb. com/publications/anogen-deep-anomaly- generator/(accessed on 20 August 2018), 2018. 112 [136] M. Ali, PyCaret: An open source, low-code machine learning library in Python, July 2020, pyCaret version 2.1. [Online]. Available: https://www.pycaret.org [137] P. Geurts, D. Ernst, and L. Wehenkel, “Extremely randomized trees,” Machine learning, vol. 63, no. 1, pp. 3–42, 2006. [138] G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y . Liu, “Lightgbm: A highly efficient gradient boosting decision tree,” in Advances in neural information process- ing systems, 2017, pp. 3146–3154. [139] S. Haykin and N. Network, “A comprehensive foundation,” Neural networks, vol. 2, no. 2004, p. 41, 2004. [140] C. M. Bishop et al., Neural networks for pattern recognition. Oxford university press, 1995. [141] A. Buffi, A. Michel, P. Nepa, and B. Tellini, “Rssi measurements for rfid tag classification in smart storage systems,” IEEE Transactions on Instrumentation and Measurement, vol. 67, no. 4, pp. 894–904, 2018. [142] R. Want, “An introduction to rfid technology,” IEEE pervasive computing, vol. 5, no. 1, pp. 25–33, 2006. [143] K. Finkenzeller, “Battery powered tags for iso/iec 14443, actively emulating load modula- tion,” in RFID SysTech 2011 7th European Workshop on Smart Objects: Systems, Technolo- gies and Applications. VDE, 2011, pp. 1–8. [144] V . Chawla and D. S. Ha, “An overview of passive rfid,” IEEE Communications Magazine, vol. 45, no. 9, pp. 11–17, 2007. [145] M. Bouet and A. L. Dos Santos, “Rfid tags: Positioning principles and localization tech- niques,” in 2008 1st IFIP Wireless Days. Ieee, 2008, pp. 1–5. [146] C. Li, L. Mo, and D. Zhang, “Review on uhf rfid localization methods,” IEEE Journal of Radio Frequency Identification, vol. 3, no. 4, pp. 205–215, 2019. [147] J. Zhou, H. Zhang, and L. Mo, “Two-dimension localization of passive rfid tags using aoa es- timation,” in 2011 IEEE International Instrumentation and Measurement Technology Con- ference. IEEE, 2011, pp. 1–5. [148] D. Vinko, T. ˇ Svedek, and M. Herceg, “Effects of power consumption and modulation of the passive rfid tag on the transmission range of backscattered signal,” in 2009 International Symposium ELMAR. IEEE, 2009, pp. 263–266. [149] J. Park and P. P. Mercier, “A sub-10-pj/bit 5-mb/s magnetic human body communication transceiver,” IEEE Journal of Solid-State Circuits, vol. 54, no. 11, pp. 3031–3042, 2019. 113 114

Abstract (if available)

Abstract A wireless body area network (WBAN) is a wireless network of wearable or implanted sensors with different medical and non-medical applications. One of these applications is in human activity recognition. Recognizing human physical activities using sensor networks has attracted significant research interest in pervasive computing due to its broad range of applications, such as rehabilitation, athletics, and senior monitoring. However, there are critical challenges inherent in the sensor-based activity recognition method, such as wireless communication technology, power consumption, coverage, and reliability. Magnetic induction (MI) is a new physical layer technique, which can address some of these problems associated with wireless communication. Moreover, a proper classification method can tackle the challenges and constraints associated with the detection stage. ❧ We begin by investigating the development of the magnetic induction technique as a new physical layer for sensor networks working around the human body. To facilitate performance analysis of the MI wireless system around the human body, we introduce a circuit model describing power transmission between the transceivers. This model exploits an analytical representation of the mutual inductance between two coils defined as a function of operating frequency, and coils’ geometry, relative position, and orientation. It provides an accurate and fast method of calculating mutual induction, a key parameter in the optimal design of an MI communication system. We then verify the model via experimental measurements and numerical simulations and investigate the effect of impedance matching on system efficiency. The validation results are encouraging and indicate that for frequencies up to 30 MHz, the human body does not affect the performance of MI communication system. We also explore the capability of the model in synthesizing time-series data using experimental measurements and comparison with ground truth data. ❧ Next, we explore the application of the MI system in human activity recognition. Critical challenges are inherent in designing a wireless network operating around a lossy medium such as the human body to gain a trade-off between power consumption, cost, computational complexity, and accuracy are described. We propose an innovative wireless system based on magnetic induction for human activity recognition to tackle these challenges and constraints. The proposed framework consists of two stages: data acquisition and detection. For the first stage, we employ the MI system model to synthesize MI motion data corresponding to several physical activities. In the second stage, machine learning techniques and deep recurrent neural networks are applied to the synthetic data to classify human movements. We present that MI signals are informative descriptors for the motion of human body parts, and therefore this approach can successfully identify human movements. ❧ Finally, we extend the magnetic induction-based human activity recognition (MI-HAR) system and study the application of the MI system in motion tracking instead of identifying the user’s activity. Analyzing body movement in 3D space facilitates behavior recognition in different applications, including healthcare, rehabilitation, sports, and senior monitoring. We propose a motion tracking system based on MI to address the issues and limitations of existing motion tracking systems based on wearable sensors. We integrate a realistic prototype of an MI sensor with machine learning techniques and investigate one-sensor and two-sensors configuration setups for motion reconstruction. The approach is then evaluated using measurements and synthesized datasets generated by the analytical model of the MI system and laboratory measurements. We show that the system has an acceptable dynamic spatial resolution compared to the ground truth variation captured by Kinect.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Theory and design of magnetic induction-based wireless underground sensor networks

PDF

Multidimensional characterization of propagation channels for next-generation wireless and localization systems

PDF

Human motion data analysis and compression using graph based techniques

PDF

Models and algorithms for energy efficient wireless sensor networks

PDF

Multichannel data collection for throughput maximization in wireless sensor networks

PDF

Deep learning for subsurface characterization and forecasting

PDF

Physics-based models and microwave sensors for mapping soil carbon in arctic permafrost using long-wavelength radars

PDF

Application-driven compressed sensing

PDF

Functional connectivity analysis and network identification in the human brain

PDF

Learning, adaptation and control to enhance wireless network performance

PDF

Learning and control for wireless networks via graph signal processing

PDF

Novel multi-stage and CTLS-based model updating methods and real-time neural network-based semiactive model predictive control algorithms

PDF

Nonlinear dynamical modeling of single neurons and its application to analysis of long-term potentiation (LTP)

PDF

Representation, classification and information fusion for robust and efficient multimodal human states recognition

Asset Metadata

Creator Golestani, Negar (author)

Core Title Magnetic induction-based wireless body area network and its application toward human motion tracking

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Electrical Engineering

Publication Date 03/18/2021

Defense Date 12/22/2019

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag analytical model,circuit model,classification,coils,data analysis,data models,deep learning,har,human activity recognition,impedance matching,machine earning,magnetic induction,MI,motion tracking,mutual coupling,neural network,OAI-PMH Harvest,preprocessing,recurrent neural network,regression,RNN,WBAN,wearable,wireless body area network,wireless communication,wireless sensor network,WSN

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Moghaddam, Mahta (committee chair), Lazzi, Gianluca (committee member), Sideris, Constantine (committee member), Ver Steeg, Greg (committee member)

Creator Email golestani.negar@gmail.com,ngolesta@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-429265

Unique identifier UC11667173

Identifier etd-GolestaniN-9330.pdf (filename),usctheses-c89-429265 (legacy record id)

Legacy Identifier etd-GolestaniN-9330.pdf

Dmrecord 429265

Document Type Dissertation

Rights Golestani, Negar

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA