Structural segmentation of point clouds with varying density based on multi-size supervoxels

Abstract

Ground objects can be regarded as a combination of structures of different geometries. Generally, the structural geometries can be grouped into linear, planar and scatter shapes. A good segmentation of objects into different structures can help to interpret the scanned scenes and provide essential clues for subsequent semantic interpretation. This is particularly true for the terrestrial static and mobile laser scanning data, where the geometric structures of objects are presented in detail due to the close scanning distances. In consideration of the large data volume and the large variation in point density of such point clouds, this paper presents a structural segmentation method of point clouds to efficiently decompose the ground objects into different structural components based on supervoxels of multiple sizes. First, supervoxels are generated with sizes adaptive to the point density with minimum occupied points and minimum size constraints. Then, the multi-size supervoxels are clustered into different components based on a structural labelling result obtained via Markov random field. Two datasets including terrestrial and mobile laser scanning point clouds were used to evaluate the performance of the proposed method. The results indicate that the proposed method can effectively and efficiently classify the point clouds into structurally meaningful segments with overall accuracies higher than 96%, even with largely varying point density.