VoxelTrack: Multi-Person 3D Human Pose Estimation and Tracking in the Wild

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Multi-person 3D pose estimation and tracking is crucial for human-computer interaction and surveillance.
• Existing methods often struggle with wide camera baselines and occlusions due to reliance on noisy 2D pose estimates.

Purpose of the Study:

• To develop a robust method for multi-person 3D pose estimation and tracking from sparsely distributed cameras.
• To overcome limitations of previous methods by directly utilizing a 3D representation, avoiding noisy intermediate 2D estimations.

Main Methods:

• VoxelTrack discretizes 3D space into voxels and computes features by averaging inversely projected heatmaps from multiple views.
• It jointly estimates 3D poses and re-identification (Re-ID) features within this 3D voxel representation.
• The approach directly estimates and tracks 3D poses, bypassing the need for cross-view 2D pose correspondence.

Main Results:

• VoxelTrack demonstrates superior performance in multi-person 3D pose estimation and tracking.
• The method achieves state-of-the-art results on challenging datasets like Shelf, Campus, Human3.6 M, and CMU Panoptic.
• Robust tracking capabilities were shown even under severe occlusions in some camera views.

Conclusions:

• VoxelTrack offers a robust and accurate solution for multi-person 3D pose estimation and tracking in challenging scenarios.
• The 3D voxel-based representation effectively handles wide camera baselines and occlusions.
• This approach significantly advances the state-of-the-art in multi-view 3D human pose analysis.