Stabilizing Skeletal Pose Estimation using mmWave Radar via Dynamic Model and Filtering

In this paper, we illustrate a method to stabilize the position estimation of human skeleton using mmWave radar. In our previous study, an optimized CNN architecture was used to extract the positions of human skeleton accurately. However, the position estimation of the joints vibrates over time. In the field of digital signal processing, filters are used to remove unwanted parts of signal and widely applied in noise reduction, radar, audio and video processing, etc. In this paper, three types of filters i.e. Elliptic, Savitzky-Golay, and Whittaker-Eilers are discussed and applied to both positions and angles of the human skeleton. This paper further presents a humanoid robotics dynamic model, specifically forward kinematics, to recalculate joint positions with improved stability. We define the root joint, a world coordinate system, and 'T' pose, to get the subsequent joints' rotation matrix using kinematics chain of the skeleton, then compute the Euler angles. After the filtering, we compare the effect of different filters using a method of Standard Deviation (SD) of the angle slope. In addition, we analyze the change of localization accuracy after recalculating the positions using forward kinematics based on the current angle, root position, and bone length information. The data collection and experimental evaluation have shown a motion stability improvement of 54.05% compared to the CNN predicted value.