Abstract: To address the geometric heterogeneity and discrete noise between vehicle-borne and backpack cross-source point clouds in complex environments,which result in issues such as feature matching ambiguity and high computational complexity in traditional point cloud registration methods,this study proposes a registration approach based on image feature matching.The method divides the registration process into vertical alignment and horizontal alignment.Vertical alignment utilizes ground points as primitives to correct deviations in the vertical direction.For horizontal alignment,the point clouds are converted into binary images through top-down orthographic projection.Potential correspondences are then constructed using Lowe's ratio test combined with gradient orientation constraints.An optimization framework based on iteratively reweighted least squares (IRLS) is developed by integrating the Huber robust kernel function to precisely estimate the parameters of the image affine transformation.Coarse registration is achieved by combining ground-based alignment with image feature matching parameters,followed by fine registration using the iterative closest point (ICP) algorithm.Experimental results on multiple real world datasets demonstrate that the proposed method effectively overcomes challenges such as point density variation in cross-source point clouds and achieves accurate registration.The mean rotation and translation errors are below 0.000 5 rad and 0.065 m,respectively,and the average per point error is less than 0.06 m.Furthermore,with the support of voxel downsampling,the method significantly improves the efficiency of large scale point cloud registration while maintaining high accuracy,showing strong potential for practical application.

Key words: point cloud registration, laser scanning, vehicle-borne and backpack point clouds, image feature matching