Feature extraction of point clouds using the DBSCAN clustering

The existing feature extraction methods often depend on the global fixed thresholds and the local information of features, resulting in sensitivity to significance of features and failure in models with different surfaces. To overcome those problems, a novel method based on DBSCAN Clustering is proposed. First, a new reverse k nearest neighbors(kNN) of points are defined as a new feature detection operator. Second, the scales of the reverse k nearest neighbors of points are utilized as the density information of points and then the introduction of the global constraints information is proposed. Finally, based on the redefinition of the concepts of the DBSCAN clustering method and the creation of a new feature recognition criterion, an improved version of the DBSCAN clustering method is used to extract features. Experimental results show that the method is simple, effective and robust, which takes into account the local information and global constraint information and outperforms existing feature detection methods on point clouds with surfaces that have diverse geometries.