Abstract: Aimed at the problems of missing detection and blocking of targets in crowded scenes in multi-target tracking, the non-local attention mechanism is introduced on the basis of the CenterTrack framework to capture the non-local dependencies between multiple targets and scenes. A tracking model based on spatial non-local attention residual block was proposed. We further extended it to the spatiotemporal domain, established a tracking model based on the spatiotemporal relationship non-local attention module, which achieved detection and tracking tasks simultaneously. The experimental results on the MOT17, MOT16 and 2DMOT15 dataset show that the two proposed online tracking models are significantly improved compared with CenterTrack, and reach the current relatively excellent level of 62.4% and 62.5% in MOTA on MOT17 challenge, which verifies the effectiveness of the proposed algorithms.