Tangential Human Motion Recognition With Micro-Doppler Signatures and One-Shot Learning

The moving direction affects the radar micro-Doppler signatures induced by human movement, with the least being attained when the human moves tangentially with regard to the line of sight. Tangential spectrograms contain minimal data on human movements and exhibit an extensive level of visual similarity, providing a significant obstacle in the classification of human motion. At the same time, the high costs associated with obtaining measured radar spectrograms result in a restricted quantity of data, making it more challenging to identify human motion spectrograms in one-shot situations. Considering the aforementioned difficulties, we propose a one-shot human motion recognition (HMR) method with tangential spectrograms based on metric learning which creatively uses simulation data as the support set and measured data as the query set. In our model, the local region enhancement (LRE) component is introduced to concentrate solely on echo signal regions near the central frequency by employing global information to improve the efficacy of valuable feature channels and diminish the impact of extraneous ones. A spatial group-wise refinement (SGR) component has been proposed to extract discriminative features from central regions by assigning weights to distinct spatial positions within each semantic group in local feature maps. Compared to other methods, the experimental results indicate that the proposed methodology enhances the classifier’s efficacy, resulting in a rise in the classification accuracy of 3.4% at least.