Micro-doppler feature and image based human activity classification with FMCW radar

Since FMCW radar measures the range, the Doppler and the angle of the targets frame by frame in time sequences, invariant micro-Doppler features become a tough challenge for deep learning based classification and recognition in radar signal processing and artificial intelligence communities. This paper presents an approach to create micro-Doppler images from range-Doppler images sequences by sum of Doppler profiles in targeted range bins. The time invariant micro-Doppler features are extracted by statistical characteristics and motion pattern of the human activities. The statistical features reflect the distribution of the micro-Doppler and the energy property of movements for different human activity classes. The motion pattern features are decomposed by Singular Value Decomposition (SVD). Furthermore, the micro-Doppler features are analyzed to find significance by Support Vector Machine (SVM). Based on these micro-Doppler significance, Faster RCNN is trained and used for detection of human activities. Experimental results show that the proposed micro-Doppler features can distinguish the human motion pattern very well and the accuracy of the Faster RCNN based human activity classification achieves more than 95%.