3D Multi-object Tracking in Autonomous Driving: A survey
Zhe Wei Peng, Wei Liu, Zuotao Ning, Qixi Zhao, Shuai Cheng, Jun da Hu · 2024
3D multi-object tracking(3D MOT) is an indispensable component of autonomous driving because of its ability to perceive and track surrounding objects. In order to track the same object in consecutive frames, 3D MOT faces challenges such as object occlusion, abrupt motion, lighting variations and distortions, as well as scenarios with small and dense objects. In recent years, various methods have been proposed to address these issues. Therefore, it is crucial to analyze the commonalities and differences among different methods. Based on the relationship between detection and data association, this paper analyzes some mainstream methods in the recent three years. It categorizes them into three classes: tracking by detection(TBD), joint detection and embedding(JDE), and joint detection and tracking(JDT). Furthermore, for each method, we further divide it into several subcategories and summarize their approaches. Additionally, we compare the advantages and disadvantages of different methods and conduct a performance analysis on the nuScenes dataset.