Behavior recognition method based on MSR enhanced micro-Doppler spectra

Abstract:
In the radar automatic recognition of human targets, the extracted human micro-Doppler features are commonly used to identify human movement states or action attitudes. Affected by human posture and noise, human micro-Doppler features are sometimes weak and vague, which is difficult to extract stably and use for classification. In this paper, a new behavior recognition method for micro-Doppler spectra is proposed based on FMCW (Frequency-modulated continuous wave) radar. Firstly, MSR (Multi-Scale Retinex) algorithm is used to enhance the constructed micro-Doppler map information. Finally, a deep convolutional neural network (ResNet50) is used to extract the micro-Doppler features of the spectrum and classify six actions such as walking and drinking. The average classification accuracy of the six actions reached 88.22%, which verified that the method of using MSR to enhance the original atlas proposed in this paper is feasible and effective for human action classification.