A probability distribution-based point cloud clustering algorithm

Abstract

Point cloud is a kind of important dataset in robot navigation and environment understanding. Most clustering algorithms were designed with certain assumptions and it is difficult to find natural clustering in a point cloud which contains lots of noise and unknown number of clusters with nature shapes. In this paper, we propose a clustering algorithm based both on density and spatial distribution of point cloud to deal with point cloud clustering problem. The algorithm combines DBSCAN and robust information-theoretic clustering method. It selects density-connected points from very dense area as core points first and then calculates the value of local volume after compression (LVAC) of border points around a cluster of core points to decide whether a border point belongs to the same cluster with these core points or not. In order to deal with real dataset, we optimise the algorithm to adapt to 3D point cloud which is used for robot navigation. Separating connected objects where points have different spatial distributions but similar density is possible according to the proposed method. Experiments on real dataset validate the proposed density distributed-based clustering algorithm.