Human Motion Classification Based on Range Information with Deep Convolutional Neural Network

In this paper, we investigate the feasibility of recognizing human motions by high resolution range information using a deep convolutional neural network (DCNN). Most existing schemes for radar-based human motion classification exploited the handcrafted features extracted from the micro-doppler signature which may be weak or indistinctive in some motions. In contrast to conventional methods, we utilize the high resolution range information to classify human motions which is more robust than the micro-doppler signature, especially when the radial velocity is not obvious, and it has a better tolerance for the incident angles. A DCNN, one of the most successful deep learning algorithms, is employed to solve the classification problem without any handcrafted features. Real data of six human subjects performing seven motions were collected using an ultra-wideband (UWB) radar with only range information. The experimental results illustrate that the DCNN can achieve an accuracy rate of 95.24% on classification of human motions based on the high resolution range information.