MULTI-OBJECT TRACKING NETWORK BASED ON ATTENTION AND MULTI-LEVEL CLUE ASSOCIATION

In multi-object tracking (MOT) task, occlusion between targets can easily lead to target tracking failures and trajectory association errors. In order to solve this problem, this paper proposes a new multi-object tracking algorithm based on attention mechanism and multi-level cue association strategy. A target visibility map was generated and converted into a spatial attention map to solve the problem of occlusion between multiple targets. In the specific target branch network, channel attention was used to improve feature robustness. In addition, a multi-level clue association strategy that combines the appearance, movement, and interaction of the target object is proposed to match the correct trajectory of the current target. The experimental results on the benchmark data sets MOT16 and MOT17 show that, compared with the existing methods, the method proposed in this paper can obtain better results on multiple evaluation indicators.