Geometrical Segmentation of Multi-Shape Point Clouds Based on Adaptive Shape Prediction and Hybrid Voting RANSAC

This work proposes the use of a robust geometrical segmentation algorithm to detect inherent shapes from dense point clouds. The points are first divided into voxels based on their connectivity and normal consistency. Then, the voxels are classified into different types of shapes through a multi-scale prediction algorithm and multiple shapes including spheres, cylinders, and cones are extracted. Next, a hybrid voting RANSAC algorithm is adopted to separate the point clouds into corresponding segments. The point-shape distance, normal difference, and voxel size are all considered as weight terms when evaluating the proposed shape. Robust voxels are weighted as a whole to ensure efficiency, while single points are considered to achieve the best performance in the disputed region. Finally, graph-cut-based optimization is adopted to deal with the competition among different segments. Experimental results and comparisons indicate that the proposed method can generate reliable segmentation results and provide the best performance compared to the benchmark methods.

Automated summary

This work proposes a robust geometrical segmentation algorithm for detecting inherent shapes from dense point clouds. The algorithm divides the points into voxels based on connectivity and normal consistency, classifies the voxels into different shapes, and extracts multiple shapes including spheres, cylinders, and cones. A hybrid voting RANSAC algorithm is then used to separate the point clouds into corresponding segments. The proposed method also considers point-shape distance, normal difference, and voxel size as weight terms for evaluating the proposed shape. Graph-cut-based optimization is adopted to handle competition among different segments. Experimental results show that the proposed method achieves reliable segmentation results and outperforms benchmark methods.