CNN Based Joint Positioning and Pose Recognition of Concealed Human for 3D Through-Wall Imaging Radar

Abstract

Information about concealed human targets behind wall including target position, pose, etc. is crucial for the through-wall sensing applications. In this paper, we propose a convolutional neural network (CNN) based through-wall multi-task processing framework, using a low-frequency 3D ultra-wideband (UWB) multiple-input-multiple-output (MIMO) radar. First, we use the 3D UWB MIMO radar to capture the concealed human echoes at different ranges and under different motions including standing, sitting, squatting and lying, with back-projection algorithm adopted to construct the 3D radar images. We label the obtained 3D radar images according to the human target motions, and divide them into the training and validation datasets at the ratio of 3:1 randomly. Then, we design a mini end-to-end full convolutional consisted of 2 modules for 3 tasks: human detection, positioning and pose recognition. One is convolutional module to extract the features of the azimuth and distance for the first two tasks, the other is residual module to extract height features for the last task, with different features contacted to improve performance of our network. The loss function of each task are: binary cross entropy, mean square error and cross entropy, with the moment estimation (Adam) as an optimizer. Finally, we get the confidence, normalized coordinates and pose category of human targets from the network. Field experimental shows that the proposed method can be used for the joint positioning and pose recognition multi-task application with high accuracy.