动态环境下具有旋转和平移不变性的激光雷达点云描述子

Lidar Point Cloud Descriptor with Rotation and Translation Invariance in Dynamic Environment

针对Intensity ScanContext(ISC)等激光雷达闭环检测算法中描述子易受载体旋转和平移变化干扰、不变性较弱从而导致闭环效果差的问题,提出了一种在城市动态环境下具有旋转和平移不变性的激光雷达点云描述子。首先,在点云处理部分,基于角度图像的地面点分割算法去除点云数据中的地面点,并利用基于弯曲体素聚类的动态目标剔除算法完成点云分割和动态目标的剔除。然后,基于惯性测量元件采集的四元数对点云进行坐标系变换,将所有帧点云统一至同一航向角下,以实现描述子的旋转不变性。接着,通过点云中静态目标与载体运动方向所在法平面的相对位置关系来确定描述子绘制中心,以实现描述子的平移不变性。最后,利用所提描述子进行闭环检测,并结合点云数据的空间结构对闭环检测结果进行一致性验证。在KITTI公开数据集下的实验结果表明,所提方法在城市动态环境下能够更加快速准确地实现闭环检测,相比于ISC算法,所提方法在准确率为100%时召回率提高了8.58个百分点,且平均耗时减少了12.90%。

Descriptors in lidar closed-loop detection algorithms such as Intensity ScanContext(ISC)are easily disturbed by carrier rotation and translation variations,and their invariance is weak,which leads to poor closed-loop effect.In view of these problems,a lidar point cloud descriptor with rotation and translation invariance in an urban dynamic environment is proposed.Firstly,in terms of point cloud processing,a ground point segmentation algorithm based on angular images is used to remove ground points in point cloud data,and a dynamic target elimination algorithm based on curved voxel clustering is adopted to realize point cloud segmentation and remove dynamic targets.Secondly,Coordinate system transformation of point clouds are carried out based on quaternions collected by an inertial measurement unit,and the point clouds in all frames are unified under the same heading angle,so as to realize the rotation invariance of the descriptor.In addition,the relative position relationship between the static targets in the point clouds and the normal plane where the moving direction of the carrier is located is used to determine the rendering center of the descriptor,so as to realize the translation invariance of the descriptor.Finally,the proposed descriptor is used for closed-loop detection,and the consistency of the closed-loop detection results is verified according to the spatial structure of the point cloud data.The experimental results under the KITTI public dataset show that the proposed method can achieve closed-loop detection more quickly and accurately in an urban dynamic environment.Compared with that of the ISC algorithm,the recall rate of the proposed method is improved by 8.58 percentage points at an accuracy of 100%,and the average time consumption is reduced by 12.90%.