A Comparative Study of Deep Learning Robustness for Sensor-based Human Activity Recognition
Sakorn Mekruksavanich, Anuchit Jitpattanakul · 2023
Recognition of human activities has become a highly significant area of research in the current era of the Internet of Things and A.I. technologies. Human activity recognition (HAR) plays a crucial role in monitoring and analyzing human behaviors using various types of sensors. Such sensor-based HAR systems are being used in various industry domains, including healthcare, professional sports, and worker assessment in intelligent factories. The integration of micro electronics mechanical systems sensor technology in smart wearable devices has significantly improved the capabilities of HAR using machine learning and deep learning techniques. However, processing longtime dependent sequence samples with noisy data is still a significant challenge, affecting the classification time and accuracy. This study aimed to analyze noise-tolerant deep learning models for recognizing physical activities from noisy smartphone data. The study analyzed the recognition performances of the five most popular deep learning networks for HAR under different levels of Gaussian noise. The results indicated that the BiGRU network demonstrated outstanding robustness to Gaussian noise of different levels.