Unveiling the Power of Self-Supervision for Multi-View Multi-Human Association and Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multi-view multi-human association and tracking (MvMHAT) is crucial for video surveillance.
• Existing methods often focus solely on tracking within a single view or across time, lacking cross-view association.
• MvMHAT requires complex annotations and offers richer data for self-learning.

Purpose of the Study:

• To develop an end-to-end neural network for MvMHAT using self-supervised learning.
• To leverage spatial-temporal self-consistency properties (reflexivity, symmetry, transitivity) for improved tracking and association.
• To introduce new large-scale benchmarks for MvMHAT research.

Main Methods:

• An end-to-end neural network architecture is proposed for MvMHAT.
• Self-supervised learning losses are designed based on symmetry and transitivity properties.
• The method optimizes appearance feature learning and assignment matrices for cross-view and temporal association.

Main Results:

• The proposed method demonstrates effectiveness in multi-view multi-human association and tracking.
• Extensive experiments on new benchmarks validate the algorithm's performance.
• The developed benchmarks and code are publicly released to foster further research.

Conclusions:

• The self-supervised approach effectively addresses the MvMHAT problem.
• The introduced benchmarks will accelerate progress in the field.
• The public release of resources supports the research community.