Feature extraction of point clouds based on region clustering segmentation

This paper proposes a feature extraction method for scattered point clouds. First, a clustering algorithm is used to divide point clouds into different regions that represent the original features. In each sub-region, we calculate the angles between the directed line segments from sampling points to the neighborhood points and set the angle threshold to identify edge feature points of uniform distribution. For the edge points of non-uniform distribution, we introduce the local neighborhood size as a discrete scale parameter for edge point detection, and then accurately identify and record the detected edge points. Then, according to the mean curvature of point clouds, the local feature weights of sampling points in the sub-region are calculated so that potential sharp feature points in a local area are detected. Finally, a minimum spanning tree of feature points is established to construct connected regions and generate feature point sets. A Bidirectional Principal Component Analysis (BD-PCA) search method is used to trim and break the small branches and multiline segments to generate feature curves. We carried out experiments on point cloud models with different densities to verify the effectiveness and superiority of our method. Results show that the edge features and sharp features are effectively extracted, and our method is not affected by the noise, neighborhood scale, or quality of sampling.