3D Human Pose Estimation and Tracking based on Feedback Mechanism

To address the challenges of 3D human pose reconstruction failures and multi-person pose tracking in natural scenes, the study introduces a feedback mechanism and pose tracking function based on mvpose model matching mechanism. In this study, we employ the extended Kalman filter to predict the current moment’s 3D pose hypothesis; this 3D pose hypothesis is aligned with the reconstructed 3D pose from the previous moment, ensuring temporal continuity and consistency. The 3D pose hypothesis is projected onto the camera view to enhance the model's performance in complex environments to achieve 2D pose fusion. In addition, it makes full use of the spatial information generated by the feedback mechanism and the re-id features matched across views to achieve 3D pose tracking and re-identification effectively. Experimental results show that the method effectively reduces the failure of 3D pose reconstruction caused by multi-person interaction scenarios and achieves 3D pose tracking and re-identification functions.