Human Action Recognition Based on Local Micro-Doppler Time-Frequency Map Sequences

The radar-based human action recognition algorithm usually employs themicro-Doppler time-frequency maps of the overall action, ignoring the fine-grained sequence features in the process. To address this problem, we propose a local micro-Doppler based action recognition algorithm, which processes the micro-Doppler time-frequency map of complete action into a local micro-Doppler time-frequency map of multiple processes. First, the distance-Doppler map is obtained by two-dimensional Fourier transformation and sliding frame of the collected intermediate frequency signal. Then, the distance-Doppler maps of consecutive frames are spliced along the velocity dimension one by one to form a multi-frame fusion micro-Doppler time-frequency map. Finally, the sequence of local micro-Doppler time-frequency maps of a complete action is input into the CNN-Transformer network for training and classification. The experimental results show that when the appropriate number of consecutive frames is selected for fusion, the average recognition accuracy of 8 kinds of human actions can reach 99.84%, and it also has good generalization performance.