Multi-Spectral-Attention Dual-Channel Network for Radar-Based Human Gait Recognition

Deep learning techniques have gained success in radar-based human gait recognition in recent years. To exploit the time-varying micro-Doppler information of human gait, most existing methods generate samples using short time Fourier transform (STFT). By extracting the features of STFT images with different sliding window lengths, dual-channel deep convolutional neural network (DC-DCNN) can obtain high accuracy for human gait. To further improve the performance of the network, this paper designs a multi-spectral-attention dual-channel network for gait recognition. Combined with the channel attention mechanism, the proposed method can achieve higher recognition accuracy without incurring additional radar resources. Experimental results have demonstrated the effectiveness of the method.