Radar-Based 3D Human Skeleton Estimation by Kinematic Constrained Learning

This paper focuses on the incorporation of kinematic constraints with the learning of 3D skeleton reconstruction without suffering from corrupted pose estimations. Specifically, we devote to exploring a kinematic constrained learning architecture that incorporates the forward kinematics constraint into building a learning model for predicting skeletal key points from observed radar data in the form of range-Doppler spectrum, which contributes to eliminating the corruption of skeleton reconstruction in the scenarios when the radar signals suffer from incomplete sensing process or insufficient granularity due to the bandwidth limitation of radar sensors. In developing our learning paradigm, we define two loss functions, namely the distance loss and the angle loss with respect to the parent-child nodes of skeleton joints, to guide the learning of the deep kinematic network, which is essential to facilitate the skeleton reconstruction without corruption. In addition, the proposed learning architecture involves a refined network module to compensate the estimation offset due to error accumulation of the kinematic model in an iterative fashion. The experimental results are presented to validate the proposed method.