Multi-object tracking algorithm based on improved TrackFormer
Shuyu Chen, X. Ren · 2023
In multi-object tracking, this paper proposes a multi-object tracking algorithm based on improved TrackFormer, aiming at the problem that the tracking object is occluded or partially occluded when multiple tracking objects interact, resulting in difficult detection of objects and wrong correlation between tracking objects. Firstly, an enhanced attention module is designed in the deformable attention mechanism used by TrackFormer, and the attention weight threshold is set on the basis of the original, and the characteristic information of the reference point itself is fused to achieve better detection effect while removing the sample point with low correlation. Secondly, the object prepositioning module is added, which can predict the possible position of the object as input to the decoder according to the characteristic information of the current frame, so as to better detect the object. The experimental results on the MOT17 dataset show that compared with TrackFormer, the improved network has a 7.8% increase in MOTA, an 11% increase in IDF1 identity score, a 13% decrease in ML loss trajectory, and a 2190 decrease in the number of ID switches, verifying its effectiveness.