CONVRNN-RESNET NETWORK FOR GAIT RECOGNITION USING MILLIMETER WAVE RADAR

Gait is a complex spatio-temporal biological feature, which is widely used for human recognition due to its characteristics of difficult camouflage and non-contact recognition. In this paper, a gait-based human identification system using a mmWave radar is proposed. The range-doppler matrix (RDM) was extracted frame by frame from the data received by the mmWave radar, and interference removal was performed afterwards. An environment-independent dataset was generated. The traditional concatenated model CNN-LSTM combines convolutional neural network (CNN) and long short-term memory network (LSTM), which cannot fully extract complex spatiotemporal features. So we proposed a model ConvRNN ResNet that combined convolutional recurrent neural network (ConvRNN) and Residual Network (ResNet) to analyze spatiotemporal sequences. Experiments show that ConvRNN-ResNet achieves up to 99.2% accuracy on the constructed testset with an observation duration of 5 s, and 94.5% accuracy under a shorter observation duration of 2.5 s, which is superior to CNN-LSTM and competent for human recognition.