The Fine segmentation of 3D laser point clouds is a challenging aspect of 3D reconstruction, often resulting in over-segmentation or under-segmentation. Addressing the complexity of vertical segmentation in point cloud buildings and the suboptimal classification outcomes, this paper presents an improved supervoxel over-segmentation algorithm that integrates normal vector angles and feature distances. Firstly, octree is used to screen seed voxels in 3D point clouds. Secondly, the similarity between seed points and non-seed points is judged by combining normal vector angle and feature distance to realize supervoxel over-segmentation, and the points with unreliable normal vectors generate in the over-segmentation process are marked. Finally cluster supervoxels semantically by using region growing algorithm, and reassign marked points by using kd tree. Experimental results show that the proposed algorithm can generate relatively complete semantic clustering of building facades, and the segmentation accuracy is 91.65% and 90.5%. Compared to the VCCS algorithm and a boundary-enhanced supervoxel over-segmentation approach, the segmentation accuracy is improved by nearly 30%.