Center point to pose: Multiple views 3D human pose estimation for multi-person.

Area of Science:

• Computer Vision
• 3D Human Pose Estimation
• Multi-Person Tracking

Background:

• Single-view 3D human pose estimation struggles with crowded and occluded scenes due to missing depth information.
• Existing multi-view methods rely on associating 2D poses, which are prone to incompleteness and noise.

Purpose of the Study:

• To develop a robust and efficient multi-view 3D human pose estimation method for crowded environments.
• To overcome limitations of 2D pose estimation and joint association in multi-person scenarios.

Main Methods:

• Proposed a Center Point to Pose (CTP) network operating directly in 3D space.
• Projected 2D joint features from multiple cameras into a unified 3D voxel space.
• CTP network regresses person centers and 3D bounding boxes, followed by detailed 3D pose estimation within each box.
• Implemented a Non-Maximum Suppression (NMS)-free approach for person center regression, enhancing efficiency.

Main Results:

• The CTP network demonstrated effective direct 3D pose estimation without relying on intermediate 2D pose association.
• Achieved competitive performance on several public datasets, validating the proposed network's efficacy.
• The NMS-free design simplified the process and improved computational efficiency.

Conclusions:

• The Center Point to Pose network offers a robust and efficient solution for multi-person 3D human pose estimation in challenging scenarios.
• Directly operating in 3D space and avoiding 2D pose limitations significantly enhances performance.
• The proposed method represents a promising advancement in computer vision for analyzing complex human interactions.