Sentinel for confidence-aware multi-object tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Multi-object tracking (MOT) struggles with detector confidence and occlusions.
• Existing methods often fail to adapt to dynamic environmental conditions.

Purpose of the Study:

• Introduce Sentinel, an uncertainty-aware MOT system.
• Enhance robustness and identity preservation in challenging tracking scenarios.

Main Methods:

• Sentinel employs Confidence Aware Association (CAA) to dynamically adjust association costs.
• Survival Boosting Mechanism (SBM) leverages weak detections to bridge occlusions.
• Per-track uncertainty diagnosis and proactive strategy optimization.

Main Results:

• Sentinel demonstrates superior performance on MOT17, MOT20, and DanceTrack datasets.
• Achieved strong Higher Order Tracking Accuracy (HOTA), Identification F1-score (IDF1), and Association Accuracy (AssA).
• Significantly improved identity preservation and reduced track fragmentation.

Conclusions:

• Sentinel offers an uncertainty-driven approach to MOT, moving beyond passive detector reliance.
• The system enhances robustness in realistic, crowded environments.
• While introducing minor overhead, Sentinel provides significant improvements in tracking accuracy and continuity.