Files
Abstract
With the proliferation of smartphones and fitness bands that have various sensors such as accelerometers, wearable sensor-based Human Activity Recognition (HAR) systems have gained wide popularity and researchers have proposed numerous techniques for recognition of these activities. Human activity recognition has many applications particularly in health care, cognitive assistance, city planning, indoor localization and tracking, and human-computer interaction. Although there has been some progress, a practical robust HAR system remains elusive because the collected data are affected by several factors such as noise, data alignment, and other constraints. In addition, the variability in the sensing equipment and their displacement is a practical challenge for implementing HAR in real-world applications.This dissertation explores the twin problems of making wearable sensor-based HAR systems robust and real time. Towards enhancing the robustness of ML-based HAR systems, we adopt feature selection methods on time and frequency domain features and apply classifiers for evaluating the recognition performance. We show the effect of different feature sets on each of the classifiers and further demonstrate in our results the impact of decreasing the size of the training set on the accuracy of the classifiers. Towards building an Online HAR system, this thesis explores the concept of Shapelets to avoid complex feature extraction. We propose a procedure to find the most representative shapelet for each activity class based on time series distance metrics and dynamic time warping. Furthermore, we generate a personalized shapelet library database driven from users' activity time series.We evaluate the proposed algorithm and techniques using a dataset comprised of accelerometer readings of 77 individuals performing various activities such as walking/jogging on treadmill, walking on different surfaces, climbing stairs, and non-ambulatory activities. Our experiments demonstrate that by using selected features from the time and frequency domain, we can achieve higher accuracy rates if we limit the training and testing sets to specific age groups. Furthermore, while we mainly use a single hip-worn accelerometer sensor as our sensing device, we show our method could support any wearable accelerometer sensor.