Enhancing Autonomous Vehicles Perception : Multi-Object Tracking With An Occlusion Handling Approach

Abstract

Our thesis presents an enhanced multi-object tracking framework built upon the OC- SORT pipeline, designed to improve robustness in challenging autonomous driving sce- narios, particularly under occlusions and missed detections. The proposed method inte- grates state-of-the-art object detectors—YOLOv8, YOLOv9, and the latest YOLOv11. Experimental evaluation was conducted on the KITTI . The system’s performance was assessed using standard tracking metrics such as HOTA, MOTA, IDF1, MT, and ML. Among all configurations, the combination of YOLOv11 with OC-SORT achieved the best results, demonstrating superior detection quality, identity preservation, and long- term tracking reliability. Specifically, it reached 61.07% HOTA, 62.51% MOTA, and 75.67% IDF1, outperforming existing state-of-the-art trackers.