Aiming at the problem of reduced tracking accuracy caused by occlusion between targets in complex scenes, an improved multi-object trcking algorithm based on the Fairmot framework is proposed. The feature map of the backbone network is used for information interaction between dimensions through a triplet attention mechanism to generate an attention mask, which improves the positioning ability of the target. Person re-identification branch adopts Circle Loss to select the degree of optimization according to the current state, to extract more accurate appearance features and distinguish different target objects. The experimental results show that the tracking accuracy on the MOT15 data set is increased to 62%, the MT (Mostly Tracked) is increased to 358, and the identity switching is reduced by 68 times. It has a better tracking effect in the scene where occlusion occurs.