Affiliations: [a] School of Computer Science, Xi’an Polytechnic University, Xi’an, China
| [b] School of Science, Xi’an Technological University, Xi’an, China
Correspondence:
[*]
Corresponding author. Shijun Xu, School of Science, Xi’an Technological University, Xi’an 710021, China. E-mail: 2934130805@qq.com.
Abstract: It is the key step of the classification and recognition to segment 3D point cloud. Aiming at the shortcomings of super-voxels concavo-convex segmentation algorithm for 3D point cloud, an efficient segmentation method based on multi-feature fusion is proposed. Firstly, the noises of 3D point cloud are removed by the statistical outlier removal filter, and the denoised 3D point cloud is simplified by the voxel grid filter. Secondly, it is divided into many voxels with the same size by the octree, and the voxels with the smallest mean curvature in local neighbourhood are used as seed voxels for regional growth to form super-voxels. Next, the super-voxels adjacency graph is structured and the concave-convex feature, continuous feature and colour feature between adjacent super-voxels are calculated as the weight on their edges. Finally, an arbitrary super-voxel is selected as the seed super-voxel to regional growth based on multi-feature weights of the edges in order to achieve the segmentation of the point cloud. The experimental results show that the proposed method whose segmentation speed, stability and accuracy are higher than existing methods greatly improves the over-segmentation or under-segmentation of the super-voxels segmentation algorithm.
Keywords: Image processing, concave-convex segmentation, 3D point cloud, multi-feature fusion, region growth, colour feature
