汽车工程 ›› 2025, Vol. 47 ›› Issue (6): 1169-1176.doi: 10.19562/j.chinasae.qcgc.2025.06.015

• • 上一篇    

基于角度交并比和自适应生命周期的三维多目标跟踪算法

孙鑫宇1,金立生1,2(),霍震1,王欢欢1,贺阳1,2,刘栋3   

  1. 1.燕山大学车辆与能源学院,秦皇岛 066004
    2.燕山大学,河北省特种运载装备重点实验室,秦皇岛 066004
    3.北京城建智控科技股份有限公司,北京 100075
  • 收稿日期:2024-10-16 修回日期:2024-12-10 出版日期:2025-06-25 发布日期:2025-06-20
  • 通讯作者: 金立生 E-mail:jinls@ysu.edu.cn
  • 基金资助:
    第二十七届中国科协年会学术论文。国家重点研发计划项目(2023YFB2504400);河北省研究生创新资助项目(CXZZBS2024064)

Three-Dimensional Multi-object Tracking Algorithm Based on Angle Intersection over Union and Adaptive Lifecycle

Xinyu Sun1,Lisheng Jin1,2(),Zhen Huo1,Huanhuan Wang1,Yang He1,2,Dong Liu3   

  1. 1.School of Vehicle and Energy,Yanshan University,Qinhuangdao 066004
    2.Yanshan University,Hebei Key Laboratory of Special Carrier Equipment,Qinhuangdao 066004
    3.Beijing Urban Construction Intelligent Control Co. ,Ltd. ,Beijing 100075
  • Received:2024-10-16 Revised:2024-12-10 Online:2025-06-25 Published:2025-06-20
  • Contact: Lisheng Jin E-mail:jinls@ysu.edu.cn

摘要:

针对智能汽车在真实交通环境跟踪周围目标时面临的轨迹错误关联和过早删除问题,本文提出一种基于角度交并比和自适应生命周期的三维多目标跟踪算法。首先,采用分数过滤与非极大值抑制分别去除低置信度和重叠的检测框,结合恒定速度模型与卡尔曼滤波消除轨迹的帧间位移。其次,考虑位置和角度因素设计角度交并比,使用匈牙利算法求解二部图最优匹配。最后,根据轨迹中断与目标距离第一性原理制定自适应生命周期策略,以动态管理轨迹状况。实验表明,改进后的方法在Waymo数据集上车辆、骑车者和行人3类目标的Mismatch依次为0.07%、0.27%和0.29%,相比于基线算法分别下降0.03%、0.18%和0.21%;在nuScenes数据集上,所提方法的IDS为312次,与基线算法相比降低49.9%。提出的角度交并比和自适应生命周期能够减少身份切换次数,改进的三维多目标跟踪算法可以获得准确稳定的时序轨迹。

关键词: 自动驾驶, 激光雷达, 三维多目标跟踪, 角度交并比, 自适应生命周期

Abstract:

For the problems of trajectory error association and premature deletion that intelligent vehicles face when tracking surrounding objects in real traffic environment, a three-dimensional multi-object tracking algorithm based on angle intersection over union and adaptive lifecycle is proposed in this paper. Firstly, low-confidence and overlapping detection boxes are removed using score filtering and non-maximum suppression, respectively, while inter-frame displacement of trajectories is eliminated by combining the constant velocity model and Kalman filter. Secondly, the angle intersection over union is designed by considering the factors of position and angle, and the optimal matching of the bipartite graph is determined through application of the Hungarian algorithm. Finally, an adaptive lifecycle strategy is formulated based on the first principle of the relationship between trajectory interruption and object distance to dynamically manage the trajectory status. The experiments show that the improved method attains 0.07%, 0.27%, and 0.29% Mismatch for Vehicle, Cyclist, and Pedestrian, respectively, on the Waymo dataset, a reduction of 0.03%, 0.18%, and 0.21% compared to the baseline algorithm. On the nuScenes dataset, the proposed method obtains an IDS of 312, a reduction of 49.9% compared to the baseline algorithm. The proposed angle intersection over union and adaptive lifecycle is able to reduce the number of identity switches, while the improved three-dimensional multi-object tracking algorithm can achieve accurate and stable temporal trajectories.

Key words: autonomous driving, LiDAR, 3D multi-object tracking, AIoU, adaptive lifecycle