基于FPFH描述子的高效点云配准与拼接方法

An Efficient Point Cloud Registration and Stitching Method Based on the FPFH Descriptor

为提高点云拼接精度,针对物体表面平整光滑而寻找关键点要求高的问题,提出1种改进的基于快速点特征直方图(FPFH)描述子的点云配准方法。对待配准的点云进行体素降采样,以此点作为关键点,并计算FPFH描述子用以描述局部特性;估计两点云之间点到点的对应关系,并经过随机采样一致性(RANSAC)算法取内点、去外点,估计初始位姿;利用迭代最近点(ICP)方法对点云进行精配准,完成点云数据的配准;依次对相邻的点云进行配准,获得旋转平移矩阵,实现点云拼接。利用公共数据集点云与自主扫描的石头点云数据,采用本文方法与传统ICP方法进行对比仿真实验,验证本文方法的有效性。结果表明:与传统ICP配准方法相比,本文方法的适应度分数为1.579E−05,显著低于点到点和点到面ICP配准方法的适应度分数;对于公共数据集不同角度下的点云拼接,本文方法的相邻点云间配准平均适应度分数达到2.058E−04,平均MSE为0.075;对于自主扫描的石头点云拼接,去除错误匹配点对后保留的对应关系基本呈现平行效果,拼接结果符合预期。使用降采样点作为关键点,并计算FPFH描述子后进行点云配准,不仅可有效降低计算复杂度,配准精度也得到显著提升,尤其在处理表面平整光滑的物体时,本文方法可为点云拼接提供良好基础,具有较好的实用性。

To improve the accuracy of point cloud stitching,an improved point cloud registration method based on the fast point feature histogram(FPFH)descriptor was proposed,addressing the challenge of high requirements for keypoint detection on smooth and flat object surfaces.The voxel down-sampling on the point clouds was performed to be registered,and these points was used as keypoints,and FPFH descriptors were calculated to describe local features.The point-to-point correspondences between the two point clouds were estimated,followed by applying the random sample consensus(RANSAC)algorithm to identify inliers,remove outliers,and to estimate the initial pose.The iterative closest point(ICP)method was utilized to achieve precise registration of the point clouds,and to complete the point cloud registration process.By sequentially registering adjacent point clouds,the rotation and translation matrices were obtained,enabling point cloud stitching.Comparative simulation experiments were conducted with public dataset point clouds and self-scanned stone point cloud data to validate the effectiveness of the proposed method against the traditional ICP method.The results demonstrate that compared to the traditional ICP registration method,the fitness score of the proposed method is 1.579E−05,significantly lower than the fitness scores of point-to-point and point-to-plane ICP registration methods.For point cloud stitching at different angles in the public dataset,the average fitness score between adjacent point clouds using the proposed method is 2.058E−04,and the average mean squared error(MSE)is 0.075.For the point cloud registration of stones scanned autonomously,the correspondences retained after removing mismatched points exhibit a nearly parallel effect,and the stitching results align with expectations.Using down-sampled points as keypoints and calculating FPFH descriptors for point cloud registration not only effectively reduces computational complexity,but also significantly improves registration accuracy.Especially when dealing with objects that have flat and smooth surfaces,the method presented in this paper can provide a good foundation for point cloud stitching,demonstrating better practicality.