부산대학교 도서관

3D multi-object tracking (MOT) is a fundamental technology in autonomous systems, playing a pivotal role across applications like autonomous driving and intelligent transportation systems. Previous 3D MOT methods mainly rely on LiDAR point clouds for object detection and tracking, often facing challenges such as occlusions and sparse data. This paper introduces SCMOT, a novel multi-modal 3D MOT framework designed to address the limitations of existing LiDAR-based 3D MOT methods. SCMOT enhances 3D object detection by filtering and refining results using semantic information, thereby reducing erroneous or redundant detections. To improve data association and enhance tracking precision, a multi-modal cost function that combines prediction confidence, semantic cues, and distance information is presented. Moreover, SCMOT can be served as a plug-and-play solution, integrating with diverse point cloud-based 3D object detectors. Extensive experiments on the KITTI tracking dataset validate the feasibility and effectiveness of SCMOT in real-world autonomous driving scenarios.