Feature-Driven Human Action Recognition from Mobile Sensor Data using Deep Learning Networks

Human activity recognition is essential for comprehending behavior patterns and forecasting the physical movements of humans, which has ramifications for domains including healthcare, sports, and security. The objective of this study was to recognize 18 primary human activities that were recorded using sensors on smartphones. This study focuses on extracting statistical features from time-domain sensory data collected by accelerometers and gyroscopes. Various deep learning techniques are utilized for activity recognition, such as 1D-Convolutional Neural Networks (1DCNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). The LSTM model achieved an accuracy of 64 percent, while the GRU model achieved an accuracy of 70 percent. The proposed 1D CNN attained a maximum accuracy of 97%. The study concludes that leveraging statistical features yields promising accuracy, with 1D CNN Network. Future research should work on fixing the problem of miss labeling in some classes to enhance accuracy.