A novel segmentation algorithm for complex 3D mesh model in computer vision

Segmentation is a critical and necessary procedure for computer vision, texture mapping and reverse engineering, which aims to digitally partition scanned point cloud or polygon mesh into subset which belongs to some kind of algebraic or NURBS surface. In this paper, a novel segmentation algorithm on triangulated boundary mesh is proposed based on boundary extraction and feature identification to extract algebraic subset, facilitating subsequent CAD model reconstruction to a great extent. Normal vector and principle curvatures are estimated firstly to identify sharp edge and strips with large curvature (fillet), which can provide a fundamental partition, allowing an accurate identification of patch features, utilizing least-square fitting and statistical error distribution. The outstanding merit of proposed method lies on the accurately hierarchical boundary identification by sharp edge and fillet independently, followed by convenient feature extraction and rather smooth subset boundary. Some complex scanned models are processed on a PC with 2.7GHZ CPU and 1G memory, with acceptable running time and satisfactory result, demonstrating the implementation and reliability of proposed method.