A Camera-LiDAR Fusion Tracking Framework with Search Area Parameters for Multi-object Tracking

Multiple Object Tracking is of paramount importance in fields like autonomous driving and robotics. This paper introduces a camera-LiDAR fusion tracking framework based on search area parameters, aiming at addressing challenges on long-range tracking and occlusions. Leveraging the unique characteristics of both sensors, we achieve multi-information fusion when the same target is detected by both sensors. Additionally, to handle temporary disappearances and prediction errors, we introduce search area parameters that gradually adapt to errors, resolving the issue of re-tracking after occlusions. On the KITTI benchmark, our approach achieves the better performance on HOTA (77.33%) and DetA (75.35%). By overcoming the limitations of both 2D MOT and 3D MOT approaches, our method enhances multi-object tracking performance, particularly in scenarios involving occlusions or distant targets.