TBA: A Tracking-By-Attention Approach to Multiple Object Tracking
Zhen-Xun Lee, Jian–Jiun Ding · 2024
Due to the development of surveillance systems, and autonomous driving, multiple object tracking (MOT) has become a critical topic in computer vision nowadays. In this paper, we propose an attention-based MOT method to well address the tracking and detection problems in dense and complex scenarios. Our approach integrates the merits of some transformer-based systems to well consider both local and semi-global information. It also adopts an hourglass backbone, a state-of-the-art DINO detector, and a multi-head attention tracker. With these techniques, long-term associations among objects can be well captured. We also incorporate a re-identification mechanism to enhance tracking capabilities after prolonged occlusions. Experimental results demonstrate that the proposed MOT method can achieve an outstanding performance, particularly excelling in handling complex and dense scenarios.