Study of Feature Fusion Models for Human Activity Recognition on Point Clouds from mmWave Radar

Human Activity Recognition (HAR) has gained significant attention as a technological response to the ever-increasing global need to monitor the health of elderly populations. This research focused on millimeter-wave (mmWave) radar technology for HAR because of its comparative advantages over competing technologies in privacy protection, superior penetration, and lack of a requirement for ambient light. This research proposes a feature fusion model for detecting activities that incorporates both voxel and Doppler information embedded in pixels. The proposed model achieves high accuracy of 95.70% for recognizing human activities, which is a significant improvement in performance over the performance using a single model, particularly in distinguishing between challenging activity categories. This paper provides an overview of the HAR system based on mmWave radar, including the design of two submodels and the composition of the feature fusion model, as well as an accuracy analysis of the proposed model through experimentation. The study also identifies several challenges that need to be addressed in future research, such as the improvement of dataset categories, the pre-processing of point cloud data, and the development of models with greater accuracy in recognizing similar activity categories.