DDMOT: Diffusion Based Multi-Object Tracking with Deep Association via Sensor Fusion

In recent years, most fusion-based multi-object tracking (MOT) works rely more heavily on LiDAR results due to the poor motion prediction performance of camera-based trackers. While these methods achieve good accuracy, they fail to fully utilize the visual feature advantages of cameras, resulting in degraded performance under occlusion scenarios. To address these limitations, we propose a diffusion model-based multi-object tracking method via sensor fusion. The diffusion model reformulates the motion prediction task as a displacement difference generation problem, which effectively handles nonlinear motion patterns. Furthermore, we design a GDIoU-based data association method and a adaptive lifecycle management system to fully leverage the perceptual capabilities of both camera and LiDAR sensors. Experimental results on the KITTI dataset demonstrate that our proposed method outperforms baseline methods in tracking accuracy.