Adaptive approach for point cloud based CAD model reconstruction

Abstract

Basic RANdom SAmple Consensus( RANSAC) approach cannot set segmentation parameters adaptively by the noise of point clouds and has no efficient way to determine whether the segmentation results are reasonable. In order to solve these problems, an adaptive approach for point cloud based CAD model reconstruction was presented. First, the approach extracted primitive shapes from point clouds by RANSAC algorithm, then it analyzed deviations of points from the fitted primitive shapes by histograms. For unreasonably segmented point cloud patches, the approach updated parameters of segmentation and repeated the primitive shape detection process. After certain rounds of iteration, the approach detected primitive shapes from point clouds reasonably. By calibrating primitive shapes' position and orientation and trimming primitive shapes according to intersection curves, the approach reconstructed the CAD model. Deviations from points to the surface of the CAD model were analyzed by error distribution graph and histogram, which demonstrated that 70. 71% of the points whose projection distance were no more than 1% of the bounding box height. The experimental results show that, by setting segmentation parameters adaptively, the approach can extract small primitive shapes from the experimental point cloud data distorted by noise with scale equal to 1% of the bounding box height.