Multi-view 3D human pose estimation based on geometric constraints

In the field of computer vision, the existing real-time multi-view 3D human pose estimation method Faster VoxelPose suffers from redundant detections and fixed plane-fusion weights in densely occluded scenarios. To address these issues, this study proposes an improved method based on geometric constraints, termed FVP-GC. During the human detection stage, a Redundancy Suppression Mechanism (RSM) is introduced. This mechanism applies geometric constraint filtering to predicted bounding boxes during training using dual thresholds of distance and overlap ratio, thereby reducing redundant detections caused by projection overlaps. During the keypoint localization stage, a Geometric Consistency Adaptive Weighting (GCAW) strategy is proposed. By evaluating the geometric consistency among orthogonal planes, the method dynamically adjusts fusion weights, improving the fusion quality of multi-plane predictions. Experimental results on the Shelf and CMU Panoptic datasets demonstrate that, while maintaining real-time inference speed, the proposed method improves pose estimation accuracy compared with Faster VoxelPose, achieving average PCP3D values of 97.7% and 97.1%, and reducing MPJPE to 17.45 mm.