Hybrid Long-Term Multi Object Tracking (HLT-MOT) System

This paper presents a Hybrid Long-Term Object Tracking (HLT-MOT) system that effectively addresses the challenges of maintaining object identities over extended periods, particularly during occlusions or temporary disappearances. Our approach combines DeepSORT's frame-to-frame tracking capabilities with a robust re-identification framework powered by DINOv2 deep learning model. The system maintains feature galleries for tracked objects and both online and offline re-identification strategies to recover lost identities. Key innovations include: (1) a dual-model approach that balances accuracy and computational efficiency, (2) GPU-accelerated batch processing for feature extraction and matching, (3) Numba-optimized spatial-temporal consistency checks, and (4) dedicated mechanisms for prioritized tracking of a primary object. Experimental results demonstrate that our system significantly outperforms traditional tracking approaches in challenging scenarios involving long-term occlusions, crowded scenes, and appearance changes, while maintaining near real-time performance. The HLT-MOT system has important applications in surveillance, autonomous navigation, and human-computer interaction domains where persistent identity tracking is essential. We open source our code with a demo video at https://github.com/rvxfahim/Hybrid_reIDHybridreID.