Abstract
In this paper, we propose a novel technique for human motion denoising by jointly optimizing kinematic and anthropometric constraints for a noisy skeleton data. Specifically, we are focused on depth-sensor-based motion capture (D-Mocap) data that are often prone to error, outliers and distortion. To capture human kinematics, we first propose a joint-level Tobit particle filter (TPF) that incorporates a unique observation model to characterize the censored measurement of D-Mocap data. A skeleton-level Differential Evolution (DE) algorithm is then integrated with the sequential Monte Carlo sampling in the TPF, allowing joint-level particles to be re-distributed and re-weighted according to the stability and consistency of skeletal bone lengths as well as the suitability of joint kinematics. This leads to an integrated TPF-DE algorithm that significantly improves the quality of D-Mocap data by making 3D joint trajectories more kinematically admissible and anthropometrically stable. Experimental results on both simulated and real-world D-Mocap show that the errors of joint positions and the bone lengths have been reduced by 30-60%, and the accuracy of joint angles has been improved by 40-60%. The proposed TPF-DE method outperforms the recent filtering-based and deep learning methods and demonstrate the synergy between the TPF and DE algorithms for effective human motion enhancement.