Multipath Feature Expansion for Detection of Human Behaviors in NLOS Region Using mmWave Radar

Abstract

The ghost echoes in radar detection of a subject behaving in a nonline-of-sight (NLOS) environment can be utilized to benefit behavior recognition. Different echoes carry unique feature information due to different multipath wave incidents and scattering directions in NLOS radar detection. By fusing the ghost echo information, the recognition of subject postures behaving in the NLOS region can be enhanced. To suppress the effects of dynamic multipath noise and ensure feature extraction from as many echoes as possible, a denoising algorithm is proposed based on frequency segregation and probability estimation (FSaPE) of the time-frequency (TF) images of human behavior. To fuse the features extracted from many echoes, a multipath-based multistage input convolutional neural network (MBMI-CNN) is proposed and trained. The scheme is demonstrated by detecting people behaving behind an L-shaped corner with 77-GHz linear frequency-modulated continuous wave (FMCW) radar. It is shown that six typical postures behaving behind the corner can be successfully classified, with an average classification accuracy of 99.17% for all the postures.