Human Action Recognition Using Dual-Stream Neural Networks with FMCW Radar

This paper presents a approach to human action recognition utilizing FMCW radar. We propose a method that extracts range-time and velocity-time features from radar echoes and processes them using a dual-stream neural network for action classification. The improved Constant False Alarm Rate (CFAR) detection algorithm, combined with a Shape Profile Constraint - Feature Map (SPC-FM) generation method, significantly enhances feature extraction by removing noise. Experimental results demonstrate an average classification accuracy of 96.4% across six common actions, showcasing the method's robustness for applications in various domains, including medical monitoring and autonomous vehicle systems.