Human Activity Recognition using Temporal 3DCNN based on FMCW Radar

Abstract

In recent years, radar-based human activity recognition has become one of the research hotspots in society, and the rapid development of deep learning also makes it widely used in this field. This paper proposes a temporal three-dimension Convolution Neural Network (3DCNN) for a comprehensive analysis of multi-domain features including time, range, Doppler and RCS. 3DCNN was designed to deal with a series of range-Doppler maps which is denoted as dynamic range-Doppler frames. Furthermore, temporal attention module is added to emphasize the sequenced relation between each frame. Extensive experiments were conducted to demonstrate its feasibility and superiority with an average accuracy rate of 95.6% in the classification of six typical daily human activities.