An improved unified online multi-object tracking algorithm combined with attention mechanism

At present, the mainstream paradigm of multi-target tracking is still tracking-by-detection, which includes two parts: the detector for locating the target and the appearance embedding model for data association. Most methods implement the two modules separately, without considering the relationship between them. However, the biggest problem of this two-stage methods is the large amount of calculation, leading to slow running speed. In this paper, we build a unified online multi-object tracking system. By integrating the object detector and the apparent embedding model into the same shared model, we can get the bounding box and the embedding model simultaneously, so as to reduce the network complexity and speed up the operation. To further improve the performance of the detector, we add an attention mechanism to weight each dimension of the output channel, so as to highlight the important foreground information and ignore the influence of the background as much as possible. The experimental results demonstrate that we can achieve competitive results on MOT16 dataset, and the best trade-off between accuracy and speed.