Multi-Human Activity Recognition Based on Sequential 4D Point Clouds Using Frequency-Modulated Continuous Wave Radar

Human Activity Recognition (HAR) is a critical research area with applications in healthcare, security, and human-computer interaction (HCI). In this study, we present a novel HAR method with multi-target human activity data collected from different positions using 4D automotive Frequency Modulated Continuous Wave (FMCW) radar. By analyzing the temporal-frequency characteristics of the radar's intermediate frequency signals and performing matrix transformation, we obtain sparse point clouds matrix data with five dimensions. Subsequently, we conduct denoising and clustering on the point clouds matrix, and segment multiple human activity into a refined point clouds matrix with a single human activity, and aggregate dispersed points to reduce the point clouds. Furthermore, we design a novel lightweight neural network tailored to the sequential 4D-radar point clouds for feature extraction and classification of the human activity. Finally, we carry out extensive experiments, and the results prove the effectiveness of the presented HAR method.