Vehicle Multi-Target Tracking Algorithm Based on Improved CenterTrack

With the rapid increase in the number of vehicles worldwide in recent years, multi-vehicle tracking has become a critical and challenging research topic in smart transportation. Although many researchers have constructed classical multi-object tracking (MOT) models, these models often lose trajectories of objects in challenging traffic environments with frequent occlusions and high object density. This significantly hinders the practical deployment of such tracking systems. In order to improve tracking robustness and accuracy when faced with these challenges, we propose a new multi-vehicle tracking algorithm built upon the CenterTrack framework. The core of our work lies in three key improvements. Each overcomes a distinct weakness in existing approaches, and together they work to improve performance. First, we use the Wise Intersection over Union (WIoU) loss to guide the model optimization and reduce the impact of label noise during training, which results in better convergence. Second, we apply an attention mechanism to reduce feature interference caused by multiple inputs (current frame, previous frame, and heatmap) of the model. Third, we employ the Sigmoid Linear Unit (SiLU) in the backbone network to further improve nonlinear feature representation. We evaluate our algorithm on the standard KITTI multi-object tracking benchmark. Experimental results show that our method achieves a Multi-Object Tracking Accuracy (MOTA) of 65.94% and an Identification F1 (IDF1) of 84.01%. This represents an improvement of 2.93% in MOTA and 2.36% in IDF1 over the baseline method. The results demonstrate not only the effectiveness of our method but also its practical usefulness and robustness in challenging road environments.