Online Multi-target Pedestrian Tracking Based on Spatiotemporal Estimation

In online multi-target pedestrian tracking, target occlusion presents a critical challenge, severely degrading performance by compromising appearance cues essential for data association in traditional methods. To address this problem, this paper proposes an online tracking model that integrates spatiotemporal information. Specifically, a motion pattern mining network (MPMNet) based on an improved \(\gamma \) -factor activation function is first designed, and the long short-term memory network is used to mine the long-term motion law of the target, significantly improving the robustness of position prediction. Secondly, a spatial coordinate estimation network (SCENet) based on the graph attention mechanism is proposed to adaptively model the spatial relationship between targets and effectively predict the coordinates of occluded targets. Finally, experimental verification on the MOT15 and MOT17 datasets shows that the method can significantly reduce the target loss rate (ML index optimization of about 23%) and the number of identity switches (IDS reduction of about 19%) in occlusion scenarios, while taking into account the real-time requirements (processing speed of 15.3 fps).