LAUSR

Accurate and real-time tracking of multiple moving objects in 3-D space is critical for intelligent transportation applications such as autonomous driving and traffic monitoring. However, tracking performance is often hindered by difficulties such as occlusion or distant objects. To this end, we present a novel 3-D-light detection and ranging (LiDAR) multiobject tracking (MOT) system based on factor graph optimization (FGO) that simultaneously tracks multiple objects and reconstructs their models. First, we develop a novel object dynamic motion model that combines prediction confidence to guide object bounding box (BB) association, enabling robust tracking for temporarily missed objects within a sliding window. Subsequently, our system takes advantage of hybrid feature measurements to form feature point factors, which are optimized together with motion model factors and BB factors to improve pose estimation accuracy. Furthermore, we create continuous 4-D object models on the timeline for better hybrid feature association by combining their optimized trajectories. Last, the MOT performance of the proposed system is evaluated on the KITTI tracking datasets. The extensive quantitative and qualitative experiments demonstrate that our system outperforms state-of-the-art methods in both 3-D MOT and classification of events, activities and relationships (CLEAR) MOT accuracy (MOTA) metrics, achieving about decimeter-level (0.20 m) relative positioning accuracy while running at 31.6 FPS.