Dynamic Graph Tracker for Multi-Object Tracking

Multi-object tracking (MOT) plays a crucial role in computer vision, particularly in applications such as autonomous driving and intelligent video surveillance. Existing MOT methods typically extract distinguishable object features, calculate an affinity matrix, and apply matching algorithms to perform trajectory assignment. However, these approaches often fail to capture the dynamic topological relationships between objects over time. To address this limitation, we propose a dynamic graph-based MOT method (DGMOT), which is designed to model and dynamically capture the evolving relationships between objects. DGMOT constructs a dynamic graph that evolves over time to effectively represent object relationships. We integrate Gated Recurrent Unit (GRU) to dynamically update node features and employ a graph attention mechanism to improve feature interaction. These innovations enable DGMOT to capture time-varying topological characteristics effectively. As a result, trajectory stability and tracking accuracy significantly improve. DGMOT achieves a HOTA score of 74.64% and a MOTA score of 88.78% on the KITTI test set. Experimental results demonstrate that DGMOT delivers competitive performance in MOT tasks.