Tracking human-like natural motion by combining two deep recurrent neural networks with Kalman filter

The Kinect skeleton tracker can achieve considerable performance with human body tracking in a convenient and low-cost manner. However, the tracker often captures unnatural human poses, such as discontinuous and vibrational movement when self-occlusions occur. In this study, we propose an advanced post-processing method to improve the Kinect skeleton using a single Kinect sensor, in which a combination of probabilistic filtering techniques and supervised learning techniques is employed to correct unnatural tracking movements. Specifically, two deep recurrent neural networks are used to improve joint velocities, as well as joint positions produced by the Kinect skeleton tracker. Moreover, a classic Kalman filter further refines positions and velocities. In addition, we propose a novel measure to evaluate the naturalness of captured joint trajectories. We evaluated the proposed approach by comparing it to ground truth obtained using a commercial optical maker-based motion capture system.