KDD: A kernel density based descriptor for 3D point clouds. (March 2021)

• Record Type: Journal Article
• Authors: Zhang, Yuhe; Li, Chunhui; Guo, Bao; Guo, Chenhao; Zhang, Shunli
• Is Part Of: Pattern recognition. Volume 111(2021)
• Publication Date: 2021-03
• Subjects: 3D feature descriptor -- Kernel density estimation -- Point cloud registration -- KL divergence; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107691
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A novel 3D local descriptor (KDD) which achieves a satisfactory and balanced performance in terms of descriptiveness, robustness, and compactness, is proposed; furthermore, the proposed KDD is very simple since it encodes the spatial distribution of points, avoiding computing any geometric attributes and needing no rotational projection operations. The KDD is combined with different matching metrics for different datasets and the strategy for selecting different matching metrics for datasets with diverse levels of resolution qualities is provided. We apply the proposed method on a real-world dataset, i.e. the Terracotta fragment models, and the favorable results demonstrate the effectiveness of KDD and highlight the utility of the proposed method. Abstract: 3D feature description is one of the central techniques that rely on point clouds since a lot of point cloud processing techniques apply the point-to-point correspondences that are achieved via feature descriptors as input data. The feature descriptor encodes the information of the underlying surface around the feature point so as to make a local surface distinguished from another. The focus of the existing descriptors is accumulating the geometric or topological measurements into histograms or encoding the 2D images that are acquired by rotationally projecting the 3D local surfaces onto 2D planes. Histograms can hardly deal with three or more dimensional information, and the rotational projection operation does bring much unnecessary intermediate computations. To overcome these limitations, in this article, a descriptor named Kernel Density Descriptor (KDD) has been presented. One core contribution of this method is to encode the information of the whole 3D space around the feature point via kernel density estimation, and another is providing the strategy for selecting different matching metrics for datasets with diverse levels of resolution qualities. We compare KDD against several representative descriptors on publicly available datasets, the experimental results demonstrate that the KDD descriptor achieves a satisfactory and balanced performance in terms of descriptiveness, robustness, and compactness, furthermore, the comparisons validate the overall superiority of our method. The benefits and applicability on object registration and recognition and 3D object reconstruction are demonstrated by the favorable results that are obtained for both public datastes and the real-world point clouds of Terracotta fragments.