UWB-Radar-Based 3D Human Pose Estimation Using Micro-Range/Micro-Doppler Images

This paper proposes an algorithm for 3D human pose estimation based on two single-channel Ultra-wideband radars. Compared to the requirement of MIMO structure in traditional radar-imaging methods, we leverage simpler hardware devices and lower power consumption. We use two Range-Doppler images obtained by the dual-base orthogonal radars as two channels to form a RDIs sample and extract the spatial features in the sample by Convolutional Neural Network to estimate the joint coordinates of the human target. In a real indoor scene, we estimate the average error of each joint for three experimental targets to be 6.10cm. The experimental results demonstrate the effectiveness of our proposed algorithm.