融合改进SAC-IA与加权ICP的高效点云配准方法

Efficient point cloud registration method integrating improved SAC-IA and weighted ICP

摘要:

为解决传统迭代最近点（iterative closest point,ICP）算法在处理巷道等大型结构化场景时,对初始位姿敏感且易因重复性几何结构陷入局部最优的问题,本文提出一种高效的两阶段点云配准框架。在粗配准阶段,通过改进采样一致性初始配准（sample consensus initial alignment,SAC-IA）算法的关键点采样策略,从下采样后的点云中快速获取全局最优初始位姿；精配准阶段设计了一种融合欧氏距离、法向量夹角与曲率信息的多维加权ICP算法,通过构建综合几何特征的优化目标函数,有效降低了巷道平直区域的误匹配率。在斯坦福公开数据集上对改进的粗配准与精配准算法分别进行了验证,结果表明:改进后的粗配准算法相比传统SAC-IA,计算效率提升了44.1%；基于几何特征的加权ICP算法相比传统ICP,配准精度提高了23.81%,耗时降低了25.9%。 在实测巷道点云的综合配准实验中,本文提出的配准框架整体表现显著优于传统方法,在整体精度与效率方面的提升均超过20%。与主流深度学习方法相比,该框架无需依赖高性能GPU,仅使用CPU即可实现处理效率数倍的提升,并在真实巷道场景中取得了更鲁棒的配准效果。实验表明,该框架能够为大型结构化场景下的点云自动配准提供一种同时满足工业级实时性与轻量化要求的高效解决方案。

Abstract:

Traditional Iterative Closest Point (ICP) algorithms are susceptible to local optima and sensitive to initial pose when applied to large-scale structured scenes, such as underground tunnels, due to their repetitive geometric features. To address these challenges, this paper proposes an efficient two-stage point cloud registration framework. In the coarse registration stage, an improved Sample Consensus Initial Alignment (SAC-IA) algorithm with an enhanced keypoint sampling strategy is introduced to rapidly obtain a globally optimal initial pose from the downsampled point cloud. Subsequently, the fine registration stage employs a novel multi-dimensional weighted ICP algorithm. This algorithm integrates Euclidean distance, normal vector angles, and curvature information into a comprehensive geometric objective function, effectively reducing the mismatch rate in the planar regions of tunnels.The proposed components were individually validated on the Stanford public dataset. Results show that the improved coarse registration algorithm increases computational efficiency by 44.1% over the conventional SAC-IA, while the geometry-based weighted ICP algorithm improves registration accuracy by 23.81% and reduces computation time by 25.9% compared to the traditional ICP. Comprehensive experiments on real-world tunnel point clouds demonstrate that the proposed framework significantly outperforms traditional methods, improving both overall accuracy and efficiency by over 20%. Furthermore, compared to mainstream deep learning methods, this framework obviates the need for high-performance GPUs. It achieves a several-fold increase in processing efficiency on a standard CPU and delivers more robust registration results in realistic tunnel scenarios. Ultimately, this framework provides a highly efficient solution for automatic point cloud registration in large-scale structured environments, fully satisfying industrial demands for real-time processing and lightweight implementation.