Browsing by Subject "Activity Recognition"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Wearable Sensor System for Home-Based Individualized Analysis of Postural Instability in Parkinson’s Disease Patients
    (2023-07) Nouriani, Ali
    This thesis presents the design, development, and evaluation of a wearable sensor system aimed at providing home-based, individualized analysis of postural instability in movement disorder patients such as subjects with Parkinson’s Disease (PD). Past studies have shown that these patients' behaviors vary significantly in clinical settings as compared to their natural home environment. Prior methods based on fall diaries and smartphone-based mobility monitoring have previously been utilized but have been found to be of limited value. Inertial Measurement Units (IMUs) are inexpensive, wearable, and readily available on smart phones and smart watches in recent years. But these consumer-grade sensors suffer from noise and drift and are usually used in combination with other measurements such as the Global Positioning System (GPS) or cameras. However, for the purpose of everyday human body motion tracking and step length estimation, GPS is not accurate enough and does not work well indoors. Camera systems are relatively expensive, impractical to widely implement and raise many privacy concerns. Hence, this study leverages advanced estimation and artificial intelligence algorithms to solely use IMU sensors for human motion analysis. This study centers on automatic human activity recognition, accurate step length estimation, human pose estimation and finally, fall risk assessment of patients over the next one year based on one week of data collected by IMU devices in their home environments. The research starts by examining the estimation of step length and other gait variables using IMU sensors. Accurate step length estimation has a number of useful health applications, including its use in characterizing the postural instability of Parkinson’s disease patients. Three different sensor configurations are studied using sensors on the shank and/or thigh of a human subject. A nonlinear estimation problem is formulated that aims to estimate shank angle, thigh angle, bias parameters of the inertial sensors and step lengths. A nonlinear observer is designed using Lyapunov analysis and requires solving an LMI to find a stabilizing observer gain. It turns out that global stability over the entire operating region can only be obtained by using switched gains, one gain for each piecewise monotonic region of the nonlinear output function. Experimental results are presented on the performance of the nonlinear observer and compared with gold standard reference measurements from an infrared camera capture system. The observer's estimates are used to fuel state-of-the-art machine learning and deep learning models, such as Convolutional Neural Networks and Long Short-Term Memory cells (CNN-LSTM). These models enable high-accuracy activity recognition. However, deep learning algorithms typically need large training datasets to be able to generalize to rare events such as near-falls in PD patients. To address this limitation, a novel algorithm combining a high-gain nonlinear observer and transfer learning, using deep learning computer vision classification algorithms, is developed for human activity recognition. The nonlinear high-gain observer precisely estimates the attitude of the human subject's chest using data from a single Inertial Measurement Unit (IMU). The signals processed by the observer are then transformed into spectrograms to create "images" of the signals' frequency response. Deep learning computer vision algorithms, pre-trained on millions of images, are fine-tuned through transfer learning. This process illustrates how to train a robust deep learning network for activity recognition even with limited datasets. Moreover, the algorithm that employs the high gain observer is demonstrated to achieve superior performance compared to the algorithm based on just raw accelerometer and gyro signals. A different activity recognition algorithm based on the use of transfer functions to represent various daily living activities that require very limited training data is also shown to work very effectively. The thesis validates the activity recognition algorithms in real-world environments and also discusses the development of behavioral biomarkers of falls. These biomarkers offer a stronger correlation with prospective fall frequency in patients than standard clinical tests, hence improving fall prediction. The research further introduces a novel method for human pose estimation that combines a high-gain observer with deep learning and kinematic modeling, providing superior full-body joint position and body-segment angle estimations from a sparse set of IMUs attached to a few locations on the subject. Lastly, the study compares home-based motion biomarkers against standard clinical metrics in predicting future falls, revealing that home data yields a higher predictive accuracy. The outcome underscores the value of assessing patients in their natural home environment, paving the way for improved treatment and fall prevention strategies for Parkinson’s disease patients in the future.