Sentinel for confidence-aware multi-object tracking

Abstract Multi-object tracking faces two persistent challenges: managing detector confidence and preventing track loss under prolonged occlusions. We introduce Sentinel, an uncertainty-aware tracker that diagnoses per-track uncertainty online and proactively optimizes its tracking strategy. Sentinel consists of two components. Confidence Aware Association (CAA) dynamically reweighs the association-cost terms according to the current track state, enabling the effective use of low-confidence detections while suppressing identity switches. Survival Boosting Mechanism (SBM) preserves tracks at risk of disappearance by exploiting weak detection signals to bridge long occlusions, thereby reducing fragmentation and maintaining identity continuity. Evaluations on MOT17, MOT20, and DanceTrack demonstrate that Sentinel achieves strong performance in Higher Order Tracking Accuracy (HOTA), Identification F1-score (IDF1), and Association Accuracy (AssA), demonstrating its strength in identity preservation and association quality. While this design introduces modest computational overhead and may increase false positives when exploiting low-confidence detections, Sentinel improves robustness in realistic, crowded environments by moving beyond passive reliance on detector outputs to uncertainty-driven, per-track optimization.