VeTrack leverages transformer-based appearance modeling and confidence-guided multi-level association for robust...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multi-Object Tracking (MOT) is crucial for autonomous driving and surveillance.
• Challenges include identity switches due to similar appearances and occlusions in dense traffic.

Purpose of the Study:

• To propose VeTrack, a novel framework for vehicle MOT.
• To address challenges of identity switches and tracking interruptions caused by occlusions.

Main Methods:

• Developed TransformerReID module using Transformer self-attention for global context and discriminative features.
• Implemented a Multi-level Association Strategy adapting matching methods based on detection confidence.
• Integrated YOLOX detection, TransformerReID, and the association strategy for the VeTrack tracker.

Main Results:

• VeTrack achieved state-of-the-art (SOTA) performance on MOT17, MOT20, and BDD100K datasets.
• Demonstrated superior handling of target occlusion and appearance similarity.
• Achieved an MOTA score of 82.2% on MOT17 and top-three FPS performance.

Conclusions:

• VeTrack effectively reduces identity switches and improves tracking robustness in complex traffic.
• The framework shows strong generalization and potential for real-world intelligent transportation systems.