王忠立, 李文仪. 基于点云分割的运动目标跟踪与SLAM方法[J]. 机器人, 2021, 43(2): 177-192. DOI: 10.13973/j.cnki.robot.200196
引用本文: 王忠立, 李文仪. 基于点云分割的运动目标跟踪与SLAM方法[J]. 机器人, 2021, 43(2): 177-192. DOI: 10.13973/j.cnki.robot.200196
WANG Zhongli, LI Wenyi. Moving Objects Tracking and SLAM Method Based on Point Cloud Segmentation[J]. ROBOT, 2021, 43(2): 177-192. DOI: 10.13973/j.cnki.robot.200196
Citation: WANG Zhongli, LI Wenyi. Moving Objects Tracking and SLAM Method Based on Point Cloud Segmentation[J]. ROBOT, 2021, 43(2): 177-192. DOI: 10.13973/j.cnki.robot.200196

基于点云分割的运动目标跟踪与SLAM方法

Moving Objects Tracking and SLAM Method Based on Point Cloud Segmentation

  • 摘要: 现有的绝大多数同步定位与地图构建(SLAM)方法是基于静态场景假设,场景中运动目标被视为干扰,它的存在会导致定位和建图精度下降甚至失败.而运动目标检测与跟踪在很多应用中又是必须的,却在求解SLAM问题时被滤除.针对这一问题,本文提出一种融合激光雷达和惯性传感器,可同时完成SLAM和运动目标检测与跟踪的方法.首先利用惯性传感器的观测结果来补偿激光雷达扫描过程中由于自身运动引起的运动失真,在运动补偿后的点云上应用全卷积神经网络(FCNN)检测出所有可能的运动目标,并基于无迹卡尔曼滤波(UKF)实现运动目标的跟踪以及动、静目标的区分.然后利用剩下的静态背景点云部分进行数据关联和运动估计,实现定位和建图.为进一步提高建图的一致性和精度,增加了闭环检测,并基于图优化的方法实现地图和轨迹的全局优化.在开源数据集KITTI及实验平台采集的数据集上进行了大量实验验证.实验结果表明,相比于传统的SLAM方法,本文方法不仅能实现运动目标的检测与跟踪,同时可完成实时、鲁棒、低漂移的车辆位姿估计与建图,且建图精度明显优于现有其他方法.

     

    Abstract: Most of the existing SLAM (simultaneous localization and mapping) methods assume that the environment is static, and the moving objects in the scene are considered as interference, which will lead to a reduced accuracy of positioning and mapping or even failure. Although the detection and tracking of moving objects are necessary in many applications, they are ignored when solving SLAM problems. For this problem, a method combining LiDAR and IMU (inertial measurement unit) is proposed to perform SLAM and detection and tracking of moving objects simultaneously. Firstly, the motion distortion caused by LiDAR motion in the scanning process is compensated by the inertial sensor. All possible moving targets are detected by the FCNN (full convolutional neural network) based on the point cloud after motion compensation. By UKF (unscented Kalman filter), the moving targets are tracked, and the static and dynamic targets are distinguished. Then the remaining static background point clouds are applied to data association and motion estimation to realize positioning and mapping. To further improve the accuracy and consistence of mapping results, the closed-loop detection is integrated, and the global optimization of the trajectory and map is realized based on graph optimization method. Many experiments are carried out on the open dataset KITTI and the dataset collected on the self-developed experimental platform. The experimental results show that, compared with the traditional SLAM methods, the proposed method can not only effectively detect and track moving objects, but also complete vehicle pose estimation and map building in a real-time, robust, low-drift manner, and the mapping accuracy is significantly better than other existing methods.

     

/

返回文章
返回