TOWARDS ROBUST SENSOR-BASED HUMAN ACTIVITY RECOGNITION IN REAL- WORLD ENVIRONMENTS

Abstract

Human Activity Recognition (HAR) using wearable sensors has become a popular research area in recent years due to its potential applications in various fields, such as healthcare, fitness, security, and smart homes. Even though numerous HAR systems are being developed, it is still challenging to create one that can accurately identify and classify human activities in actual environments. This dissertation presents methods for recognizing human activity using a single accelerometer-based system. The research explores the two pillars of making wearable sensor-based HAR systems robust and reliable: a free-living dataset that represents real-world scenarios and a user-independent system. Towards enhancing the robustness of the sensor-based HAR system, we applied deep analysis to several machine-learning techniques and models for identifying human activity using a pseudo-free-living dataset obtained from 20 participants at the University of Georgia. We found that hierarchical meta-classifiers outperformed deep learning and classical models by 6% for classifying seven activities. We classified the metabolic equivalent (METs) levels of physical activities and achieved 80% inter-subject accuracy. We introduced model personalization, and it increased the accuracy to 87% by including 50% of the participant's data. This approach is promoted since it lowers the inter-subject variability of the dataset. We built a user-independent sensor-based human activity recognition system to explore the impact of using demographic data and anthropometric features to improve the classification of the metabolic equivalent (METs) level of physical activities based on free-living data. We used a wearable accelerometer dataset collected by 270 participants from different cities in the state of Georgia performing various physical activities. We found that including demographic data and anthropometric features in the models improves their accuracy in classifying MET levels. We built and modified a transformer's self-attention mechanism to analyze motion signals over time, which expresses individual relationships between signal levels within a time series. Furthermore, model personalization was able to reduce the dataset's inter-subject variability and raised accuracy to 94.84% by including only 30% of the participant's data in training. Achieving high performance for subject-independent systems remains challenging when using real-world data.