空间信息约束的改进ICP算法大场景点云快速配准方法

Improved ICP algorithm for fast registration of point clouds in large-scale scenarios with spatial information constraints

针对大场景点云快速配准对初值要求高的问题,将测站空间位置与三维角点特征作为空间约束信息,改进传统迭代最近邻点法(ICP)配准算法。考虑点云密度与扫描仪的距离关系,结合使用点—点准则与点—线准则进行点云精配准。实例验证结果表明,在4 507万点云场景下,均方根误差(RMSE)达到0.231 1 m,较直接使用ICP算法提高0.352 1 m,较使用采样随机采样一致性算法RANSAC+迭代最近点算法(ICP)方法提高0.119 3 m,时间分别缩短5.87、18.32 s;在843万点云场景下,RMSE达到0.051 6 m,较直接使用ICP算法提高1.052 1 m,较使用RANSAC+ICP方法提高0.266 9 m,时间分别缩短2.10、19.43 s;提取到的有效角点较Harris3D算法提高了34.84%,证明本文算法能够用于大场景散乱点云的快速配准。

In response to the high initial value requirement for rapid registration of point clouds in large-scale scenarios,the traditional iterative closest point(ICP)registration algorithm was improved by using the spatial position of the measuring station and three-dimensional(3D)corner point features as spatial constraint information.By considering the relationship between point cloud density and the distance of the scanner,point-to-point criteria and point-to-line criteria were both used for precise point cloud registration.The example verification results show that in a scenario with 45.07 million point clouds,the root mean square error(RMSE)reaches 0.2311 m,which is 0.3521 m higher than that using the ICP algorithm directly and 0.1193 m higher than that using the random sample consensus+ICP(RANSAC+ICP)method.The time is shortened by 5.87 s and 18.32 s,respectively.In the scenario with 8.43 million point clouds,the RMSE reaches 0.0516 m,which is 1.0521 m higher than that using the ICP algorithm directly and 0.2669 m higher than that using the RANSAC+ICP method.The time is shortened by 2.10 s and 19.43 s,respectively.The extracted effective corner points have increased by 34.84%compared to those by using the Harris 3D algorithm,proving that the proposed algorithm can be used for fast registration of scattered point clouds in large-scale scenarios.