摘要
针对车载激光扫描在重访道路时多趟点云的配准问题,提出一种地面点和杆状物结合的点云配准方法。配准前通过梯度算法提取地面点,依据空间相对分布关系对目标点集与待配准点集进行自动配对。考虑到传统迭代最近点(ICP)算法对初始位置要求高等局限性,采取先高程后平面的配准步骤:首先高程配准,基于地面点使用体素滤波器强化地形特征,通过利用距离约束条件来获取准确匹配点序列并计算初始配准参数,为精确配准提供良好的位姿信息;其次平面配准,以杆状物为配准基元,在利用直通滤波限定杆柱状剖面的基础上添加表面曲率特征,并设定阈值剔除错误邻近点对,从而提高配准精度和速度;最后根据线性内插实现长路线点云平滑。实测数据验证了该方法的有效性,三轴配准参数残差低于4 cm,均方根误差在3 cm左右,配准效率较高,可为大场景车载激光点云的配准提供技术参考。
Objective Vehicle-borne mobile measurement system has been widely used in many industries and departments because of its high accuracy,fast speed and rich information.Vehicle-borne point cloud also plays an increasingly important role in the task of real scene three-dimensional reconstruction.In practical applications,due to the blocking of Global Navigation Satellite System(GNSS)signal by viaducts and high-rise buildings in urban areas,the calculated revisited road point clouds have problems of layering and offset,so that they cannot meet the needs of actual engineering projects.In order to improve the quality of vehicle-borne point cloud data,it is necessary to correct the position deviation of point cloud by registration technology.At present,the registration algorithms combining deep learning and feature extraction have been widely studied,but they mainly focus on the ground fixed stations,indoor and smallscale sample point clouds.There are relatively few studies on vehicle-borne point cloud registration.The traditional registration algorithms applied to large scene vehicle-borne point clouds still have the limitations of low accuracy and low efficiency.Aiming at the above problems,a point cloud registration method combining ground points and rod objects is proposed in this paper.Methods In the proposed method,firstly,the ground point cloud is extracted based on the gradient algorithm and the elevation density distribution function.Then,the mileage segmentation is used to segment the long route point cloud to calculate the overlapping area of two point clouds by using the extreme value range of the ground point.The elevation difference is constrained to automatically generate a stable matching relationship between the target point set and the point set to be registered.Secondly,aiming at the limitation of iterative closest point(ICP)algorithm with high requirements for initial position,the registration process is divided into two steps:the elevation registration based on ground points and the plane registration based on rod objects.The elevation registration uses voxel filter method to strengthen terrain features based on ground points,obtains accurate matching point sequence and calculates initial registration parameters by using distance constraints,so as to provide good pose information for the subsequent fine registration.The plane registration takes the rod objects as the registration primitive.The surface curvature feature is added on the basis of the pass-through filter to limit the cylindrical section of the rods,and the threshold is set to eliminate the wrong adjacent point pairs to improve the registration accuracy and speed.Finally,the point cloud smoothing of the long route is realized by linear interpolation.Results and Discussions The proposed registration method is used for vehicle-borne point cloud registration by using SSW vehicle-borne mobile measurement system to collect experimental data,including those obtained on urban roads and tens of kilometers of urban expressways and highways.After ground filtering(Fig.6)and automatic matching(Fig.7)of revisited point sets,the elevation registration results(Fig.8)show that the registration method proposed in this paper can accurately register two ground point clouds with good coarse registration effect,providing robust initial pose for the plane registration.Subsequently,the improved ICP algorithm is used for plane fine registration(Fig.9).Compared with mainstream algorithms such as RANSAC-ICP and GICP(Fig.10),it is shown in Table 3 that even if the spatial distribution of the vehicle-borne point clouds in the large scenes is discrete and some ground objects are missing,the overall registration accuracy of the proposed algorithm is high,the calculation efficiency is increased by more than three times,and the high-efficiency and high-precision registration is realized.Compared with the traditional manual interactive registration results(Fig.11),the translation deviations in the X and Y directions are 0.04 m,and that in the Z direction is0.03 m.The root mean square error is about 0.03 m,which can meet the application requirements of point cloud registration.Conclusions Aiming at the problem of inconsistent position of multi-trip vehicle-borne laser point clouds on the revisited road section,we propose a fine registration method using the combination of ground points and rod objects.In this method,the rigid correspondence relationship between two point clouds is established by preprocessing such as ground point extraction,mileage segmentation and overlapping area calculation,and the registration process is divided into two stages:first elevation registration and then plane registration.Typical ground points and rod objects are used as the registration primitives.Combined with voxel filtering,spatial distance constraint and limited curvature threshold,ICP algorithm is improved to calculate the rotation matrix and translation vector.The results show that the method proposed in this paper can achieve automatic registration under the condition of complex point cloud objects,multiple noise points and no prior information,complete the high fusion of point clouds and improve the registration efficiency.Compared with the mainstream methods,facing the complex large scene urban environment,the robustness and universality of the improved ICP algorithm proposed in this paper are stronger,and the registration error is generally less than0.04 m.In a word,this method is simple and accurate in practical applications.
作者
许梦兵
刘先林
仲雪婷
张攀科
陈思耘
Xu Mengbing;Liu Xianlin;Zhong Xueting;Zhang Panke;Chen Siyun(College of Resource Environment and Tourism,Capital Normal University,Beijing 100048,China;Beijing GEOVision Tech.Co.,Ltd.,Beijing 100070,China;Chinese Academy of Surveying&Mapping,Beijing 100830,China)
出处
《中国激光》
EI
CAS
CSCD
北大核心
2023年第2期101-115,共15页
Chinese Journal of Lasers
基金
国家自然科学基金(42071444)。
关键词
图像处理
车载激光扫描
点云配准
地面滤波
杆状特征
表面曲率阈值
迭代最近点算法
image processing
vehicle-borne laser scanning
point cloud registration
ground filtering
rod features
surface curvature threshold
iterative closest point algorithm
