中心均匀化密度峰值聚类的激光点云分割

Laser point cloud segmentation based on density peak cluster of center homogenization algorithm

针对常规的密度峰值聚类算法在确定数据聚类中存在聚类中心的重复性、聚类不稳定、不适用于三维点云分割等问题,提出了中心均匀化聚类群融合算法。该算法对局部密度和距离函数进行归一化处理,较好地解决了这两种函数尺度不一的问题;基于局部密度和距离函数乘积的变化率来确定聚类中心,并对重复或距离很近的聚类中心进行了消除,避免了聚类中心非均匀分布对聚类的影响;利用数据点到聚类中心距离逐个确定每个数据的聚类归属,依据邻近聚类数据群之间的距离来判断邻近聚类之间的融合,实现对点云数据的有效分割。基于二维离散数据聚类及不同分辨率点云数据分割的实验结果表明:所提算法不仅适用于二维离散数据的聚类,也适用于三维点云数据的分割,且分割精度和稳定度要优于常规的CFDP、K-means、DBSCAN、DPC聚类算法和深度学习方法。

There are some problems in the density peak clustering,such as repeatability of cluster center,instability of clustering and not applied to the segmentation of three-dimensional point cloud.In view of this,this paper proposed a cluster group fusion of center homogenization algorithm.The algorithm normalizes the local density and distance functions and solves the problem that the two functions have different scales.Clustering centers are determined based on the change rate of the product of local density and distance function.Clustering centers with repetition or close together are eliminated to avoid the influence of non-uniform distribution of clustering center on clustering.The clustering attribution of each data is determined one by one according to the distance between points and clustering center.The fusion of adjacent clusters is judged according to the distance between adjacent clustering groups,so as to achieve the segmentation of point cloud.Experimental results based on 2 D discrete data clustering and point cloud segmentation with different resolutions indicate that the proposed algorithm is not only suitable for 2 D discrete data clustering,but also suitable for 3 D point cloud segmentation,and the segmentation accuracy and stability are better than the traditional CFDP,K-means,DBSCAN,DPC clustering algorithm and deep learning methods.