Multi-object tracking (MOT) aims to detect and track each object in a video sequence. A major challenge in MOT is occlusion. When multiple objects overlap within the field of view, their reduced visible size often leads to lower detection confidence. Existing methods typically categorize detections into high-confidence and low-confidence groups to perform two-stage matching. Given that the appearance features of low-confidence detections are unreliable, these methods typically use intersection-over-union (IoU) for similarity evaluation, followed by the Hungarian algorithm for matching in the second stage. However, relying solely on IoU can be inadequate, leading to incorrect associations. This paper proposes GMMotion, which aims to extract more effective information from low-score detections and trajectories in the second matching stage, transforming the MOT problem into a graph matching task. To better adapt graph matching for the MOT problem, our Sparse Graph Generation Model creates connections only between each pedestrian and their immediate neighbors to reduce interference edges. Our Neighbors Interaction Model and Graphs Interaction Model extract the shape and contextual information of bounding boxes as node features, and further refine these features through cross-graph interactions. By integrating GMMotion into BoT-SORT, we propose BoT-GMMotion, significantly enhancing tracking performance under occlusion conditions. This integration allows the model to better maintain accurate tracking in complex scenarios, reducing the impact of occlusions. As a result, BoT-GMMotion achieves 63.2 HOTA, 78.0 MOTA, and 77.3 IDF1 on the MOT20 test set, demonstrating its effectiveness in challenging scenes.