Robust Multi object Tracking for Autonomous Vehicles Navigation in complex Environments

Abstract

Our work presents a robust 3D multi-object tracking system designed to support au- tonomous vehicle navigation in complex and dynamic urban environments. The pro- posed approach adopts the tracking-by-detection paradigm, leveraging PointRCNN as a 3D object detector to generate high-precision detections from LiDAR point clouds. For tracking and data association, a Kalman Filter is employed for motion prediction, while the Hungarian algorithm efficiently handles matching between detections and ex- isting tracks. The system is evaluated on the challenging KITTI benchmark across three key object categories: cars, pedestrians, and cyclists. Quantitative results demonstrate superior performance compared to several state-of-the-art methods, achieving the high- est MOTA for all categories (84.81% for cars, 68.19% for pedestrians, and 83.38% for cyclists). The detection module also shows strong overall performance with a mean F1- score of 89.66%. Qualitative evaluations confirm the system’s robustness under diverse real-world conditions, including occlusions, varying lighting, and cluttered scenes.