Re-Detection Module to Mitigate Tracklet Fragmentation in Autonomous Driving

This paper focuses on the application of Detection-Based Tracking in autonomous driving environments. In such systems, object trajectories are crucial for assessing collision risks between entities such as pedestrians and vehicles. However, tracklet fragmentation caused by missed detections can undermine tracking stability and compromise safety. To address this issue, we propose a simple yet effective re-Detection module that mitigates the occurrence of tracklet fragmentation by resolving missed detections. When an object’s trajectory is interrupted in the current frame, a cropped image is generated by referencing the memory of tracklets from the previous frame, containing all affected objects. The re-Detection module then reprocesses the cropped image, fuses the new detections with those from the original frame, and removes redundant results using a Crop-Intersection over Single strategy. Finally, the recovered tracklets are updated and stored. We validate the proposed method through experiments conducted on a tram test track located in Osong-eup, Korea, evaluating various combinations of detectors and trackers. Experimental results demonstrate that the re-Detection module effectively reduces tracklet fragmentation caused by missed detections and maintains high performance even when combined with lightweight or traditional trackers. This makes it a practical and efficient solution for real-time autonomous systems.