BEVTrack: Multi-View Multi-Human Registration and Tracking in the Bird's Eye View

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Multi-view human tracking is crucial for surveillance and robotics.
• Existing methods often require camera calibration or direct bird's eye view (BEV) data, limiting real-world applicability.

Purpose of the Study:

• To address the challenge of multi-view multi-human tracking in BEV without camera calibration or explicit BEV video.
• To propose a novel scheme, BEVTrack, for robust tracking in complex scenarios.

Main Methods:

• A virtual BEV transform module generates BEV representations from individual camera views.
• A unified BEV alignment module fuses these representations using self-supervised losses for spatial and temporal consistency.
• A camera-subject collaborative registration and tracking strategy leverages inter-dependencies for enhanced accuracy.

Main Results:

• The proposed BEVTrack scheme demonstrates effectiveness in multi-view multi-human tracking.
• Experimental results on a newly built benchmark validate the method's performance.
• The approach successfully achieves tracking without requiring camera calibration or captured BEV video.

Conclusions:

• BEVTrack offers a practical solution for multi-view multi-human tracking in real-world applications.
• The method's ability to track without calibration or BEV video marks a significant advancement.
• The developed benchmark facilitates future research in this domain.