Efficient 3D Position Estimation in Badminton Scene

The badminton analysis system aims to estimate the 3D positions and subsequently provide training assistance. Despite existing 3D human pose and ball trajectory estimation methods achieving this, they are often limited to a single task. This challenge arises from two aspects: 1) Occlusion, motion blur, and similar object phenomenons. 2) The distribution of 3D spheres and humans in motion space differs significantly, making it difficult to employ a unified method. To address these issues, we propose a dual-view method that circumvents the occlusion problems associated with single-view methods while reducing the cost of multi-view methods. Specifically, we detect the 2D keypoints under dual views. Based on epipolar geometry, we propose 2D-3D uplifting module to fuse 2D keypoints to obtain accurate 3D coordinates. For 3D human pose estimation, we additionally propose a Player Tracker module based on badminton court rules to rapidly associate player IDs with bounding boxes.