一种用于道路障碍物识别的激光点云聚类算法  

Clustering algorithm of laser point cloud for road barriers detection

作　　者：张名芳[1] 刘新雨[2] 付锐[1,3] 蒋拯民 李星星[2] 

机构地区：[1]长安大学汽车学院交通运输系,陕西西安710064 [2]长安大学汽车学院车辆工程系,陕西西安710064 [3]长安大学汽车运输安全保障技术交通行业重点实验室,陕西西安710064

出　　处：《激光与红外》2017年第9期1186-1192,共7页Laser & Infrared

基　　金：国家自然科学基金项目(No.51178053;No.61374196);"十二五"国家科技支撑计划项目(No.2014BAG01B05);中央高校基本科研业务费专项资金项目(No.310822161009;No.310822171118;No.310822172001);教育部长江学者和创新团队发展计划项目(No.IRT1286);陕西省自然科学基金项目(No.2016JQ5096)资助

摘　　要：提出一种适用于道路障碍物识别检测的聚类算法,该算法用来处理各向异性分布的激光点云数据。算法的基本思想是:针对点云空间分布的实时变化,提出在线学习合并阈值的层次聚类算法,以确定聚类数搜索范围上界和初始聚类中心的待选点集;然后提出距离乘积最大化方法,对待选点集进行初始化排序,既结合点云的空间密度分布改善了聚类结果,又克服了传统K-means算法初始聚类中心难确定的问题;最后选取Silhouette和距离评价函数为聚类有效性指标分析算法的聚类效果,确定最佳聚类数。用以上自适应、在线学习的算法对2.5D激光雷达采集的点云数据进行聚类,并与其他两种聚类算法进行实际试验比较发现,本算法可以正确分割大多数空间分布各异且相互连接的障碍物。A clustering algorithm for road barrier detection is proposed, and it is used to deal with laser point cloud with anisotropic distribution. Considering the real-time change of the point cloud space distribution, a hierarchical clustering method based on online learning merging-threshold is proposed to determine the upper limit of search range of the cluster number and the selected point set of initial cluster centers. Then the method of maximizing the product of distances is proposed for initializing the point set, which improves the clustering result by incorporating spatial distri- bution of point cloud and determines the initial cluster centers of K-means algorithm effectively. Finally the optimal number of clusters is determined by the analysis of validity index. This adaptive online learning algorithm is used to cluster the 2.5D laser point cloud in the real experiments, and it can cluster most barriers with anisotropic distribution and close connection correctly in comparison with two other algorithms.

关 键 词：K-MEANS算法 激光点云 层次聚类 初始聚类中心 

分 类 号：TN958.98[电子电信—信号与信息处理] TP391.4[电子电信—信息与通信工程]