汽车工程 ›› 2025, Vol. 47 ›› Issue (9): 1700-1711.doi: 10.19562/j.chinasae.qcgc.2025.09.006

• • 上一篇    

基于场景流不确定性模型的MOT SLAM算法研究

苏家靖,朱元,陆科()   

  1. 同济大学汽车学院,上海 201804
  • 收稿日期:2025-01-23 修回日期:2025-02-22 出版日期:2025-09-25 发布日期:2025-09-19
  • 通讯作者: 陆科 E-mail:luke@tongji.edu.cn
  • 基金资助:
    江西省重点研发计划项目(20224BBE52003)

Research on MOT SLAM Algorithm Based on Scene Flow Uncertainty Model

Jiajing Su,Yuan Zhu,Ke Lu()   

  1. School of Automotive Studies,Tongji University,Shanghai 201804
  • Received:2025-01-23 Revised:2025-02-22 Online:2025-09-25 Published:2025-09-19
  • Contact: Ke Lu E-mail:luke@tongji.edu.cn

摘要:

多目标跟踪(multi-object tracking,MOT)结合同时定位与建图(simultaneous localization and mapping,SLAM)能够充分运用场景中动静态信息,可提升定位精度和鲁棒性而受到关注。本文提出了一种基于场景流不确定性模型的3D目标跟踪SLAM算法,以双目或RGB-D图像作为输入,结合实例掩膜和IMU信息,精确检测动态特征并联合估计自身和物体的位姿变换。针对动、静态和临时静态特征无法准确识别的问题,结合实例信息和场景流不确定性建模剔除误差干扰,实现精准的动态特征检测;针对运动物体特征点稀少和跟踪困难的问题,结合KLT光流和实例信息进行鲁棒的多层级数据关联;通过构建因子图并引入车辆运动学约束,实现了自身和运动物体位姿以及地图点坐标的紧耦合优化。最后,在公开数据集上进行对比实验。结果表明,所提算法能够准确跟踪自身和运动物体的位姿变换。

关键词: 视觉惯性SLAM, 3D目标跟踪, 场景流不确定性模型, 实例分割

Abstract:

Multi-object Tracking (MOT) combined with Simultaneous Localization and Mapping (SLAM) , making full use of dynamic and static information in the scene, can improve positioning accuracy and robustness, which has received considerable attention. In this paper a 3D object tracking SLAM algorithm based on the scene flow uncertainty model is proposed. With stereo or RGB-D images as input, combining instance masks and IMU information, it can accurately detect dynamic features and jointly estimate the pose transformation of itself and the objects. For the problem that dynamic, static and temporary static features cannot be accurately identified, instance information and scene flow uncertainty modeling are combined for modeling to eliminate error interference and achieve accurate dynamic feature detection. For the problem that feature points of moving objects are scarce and difficult to track, KLT optical flow and instance information are combined to perform robust multi-level data association. By constructing a factor graph and introducing vehicle kinematic constraints, tightly coupled optimization of the vehicle's own and moving object poses and map point coordinates is achieved. Finally, comparative experiments are conducted on public datasets. The results show that the proposed algorithm can accurately track the pose transformation of itself and the moving objects.

Key words: visual-inertial SLAM, 3D object tracking, scene flow uncertainty model, instance segmentation