Human Activity Recognition With FMCW Radar Using Few-Shot Learning

Frequency-modulated continuous-wave (FMCW) radar-based human activity recognition algorithms often require a large amount of sample data and have a high computational complexity. To address this problem, we propose a learning technique based on few-shot learning using less samples for FMCW radar human activity recognition. The wavelet transform and background frame differences processing on recorded human activity signals are first leveraged to obtain noise-reduced and frame-calibrated data, and then, the 2-D Fourier transform is applied to construct a range-Doppler map (RDM). Next, the maps were stitched together frame by frame along the velocity dimension to form micro-Doppler signatures (m-D signatures), which are used for feature extractions of different human activities. Finally, these m-D signatures are put into a designed residual block prototypical networks for training and classification. The experimental results show that our method is less computationally complex and more generalizable, with an average recognition accuracy of 98.33% for eight human activities with only 30 training samples.