MD-Pose: Human Pose Estimation for Single-Channel UWB Radar

Abstract

Human pose estimation based on optical sensors is difficult to solve the situation under harsh environments and shielding. In this paper, a Micro-Doppler (MD) based human pose estimation for the single-channel ultra-wideband (UWB) radar, called MD-Pose, is proposed. The MD characteristic reflects the kinematics of the human and provides a unique method for identifying the targets posture, which offers a more comprehensive perception of human posture. We explore the relationship between the human skeleton and the MD signature, which reveals the fundamental origins of these previously unexplained phenomena. The single-channel UWB radar is widely used because of its small size, low cost, and portability. In contrast, its resolution is lower than that of the MIMO UWB radar. Therefore, this paper reveals how to implement fine-grained human posture based on the MD signature with fewer channels. The MD spectrogram of the human target is obtained by the short-time Fourier transform (STFT), which is the input data of the proposed MD-Pose. A quasi-symmetric U-Net neural network is trained with the UWB radar MD spectrogram, which can estimate the human keypoints. The experiments show comparable quantitative results with the state-of-the-art human pose estimation method and provide the underlying insights needed to guide the design of radar-based human pose estimation.