融合高斯核及指数函数聚类的点云目标物提取

Point Cloud Object Extraction Based on Gaussian Kernel Function and Exponential Function Clustering

针对聚类算法的聚类中心重复性和无法对点云聚类的问题,提出了融合高斯核及指数函数的聚类中心均匀化的点云聚类方法,以优化聚类中心的均匀化分布,实现点云的均匀化聚类。首先,根据高斯核函数及密度指数函数确定局部密度,再依据局部密度的大小确定距离参数。其次,依据局部密度和距离参数的乘积确定聚类中心,同时消除聚类中心的邻近化,使得聚类中心更加均匀分布于整个数据集中。最后,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,并合并邻近聚类实现点云目标物的提取。将该算法与常规的基于密度峰值的聚类算法(clustering function based on density peak,CFDP)、K-means聚类算法、具有噪声的基于密度的聚类方法(density-based spatial clustering of applications with noise,DBSCAN)进行比较,该文所提方法可以对教室内3排椅子实现100%的提取。与相对密度关系的峰值聚类(density peak clustering,DPC)算法及深度学习方法相比,所提方法对不同分辨率目标物点云的提取精度均为96.7%,在计算效率和精度方面均优于其他两种方法。

In view of problems of repeatability of cluster center and disability of conducting point cloud clustering,a point cloud clustering method of clustering center homogenization combining Gaussian kernel and exponential function is proposed to optimize homogenization distribution of cluster centers and achieve the homogeneous clustering of point cloud.Firstly,local density is determined according to the Gaussian kernel function and density exponential function,and distance parameters are determined according to the size of local density.Then cluster centers are determined according to the product of the local density and distance parameters,and the proximity of the cluster centers is eliminated,so that the cluster centers are more evenly distributed in the entire data set.Finally,the distances between the data point and the cluster centers are used to determine the cluster attribution of each data,and the neighboring clusters are combined to achieve the extraction of the point cloud target.This algorithm is compared with the clustering function based on density peak(CFDP),K-means clustering algorithm,DBSCAN(density-based spatial clustering of applications with noise)algorithm,and the advantages of clustering algorithm in this paper are confirmed.Compared with the DPC algorithm and the deep learning method,the accuracy of objects point cloud extraction with different resolutions is 96.7%.The proposed method is prior than the other two methods in terms of computational efficiency and precision.