一种基于目标点云分布特性的动态聚类算法  

A dynamic clustering algorithm based on the point clouds distribution characteristics of obstacle

作　　者：李彩虹 何晨阳 高锋[1,2] 陈佳欣 LI Caihong;HE Chenyang;GAO Feng;CHEN Jiaxin(College of Mechanical and Vehicle Engineering,Chongqing University,Chongqing 400044,China;State Key Laboratory of Intelligent Vehicle Safety Technology,Chongqing 401120,China)

机构地区：[1]重庆大学机械与运载工程学院,重庆400044 [2]智能汽车安全技术全国重点实验室,重庆401120

出　　处：《汽车安全与节能学报》2024年第2期261-267,共7页Journal of Automotive Safety and Energy

基　　金：汽车安全与节能国家重点实验室开放基金项目(KFY2209);汽车协同创新中心揭榜挂帅项目(2022CDJDX-004);重庆市技术创新与应用发展专项(CSTB2022TIAD-KPX0139)。

摘　　要：激光雷达在自动驾驶系统的目标检测任务中发挥着重要作用,但其扫描机理会使得点云分布不均匀,常规聚类算法由于参数固定会导致较多的错误聚类。为解决该问题,该文以椭圆形状作为邻域空间,设计基于采样点位置的邻域自适应调整策略,提出一种基于目标点云分布特性的动态聚类算法。通过正确聚类、过聚类等综合结果评估算法的性能,在KITTI数据集上进行了数值分析得到算法参数,并在校园环境中进行了实车对比实验。结果表明:所提算法能减少基于密度的噪声应用空间聚类(DBSCAN)中固定邻域所造成的70.60%过聚类、49.76%欠聚类等错误结果,从而有效提高算法的综合聚类性能。The lidar sensor plays an important role in the object detection of automatic driving systems,but the spatial distribution of point cloud is uneven because of its scanning mechanism,in which case a bunch of erroneous is yielded by the conventional clustering algorithms with fixed parameters.To solve such problems,a dynamic clustering algorithm based on the distribution characteristic of object point clouds was proposed,using the elliptical shape as the spatial neighborhood which adjusted its size according to the position of the sampling points.The key parameters were further designed quantitatively with the KITTI dataset considering comprehensive clustering performances,and the comparison experiment was conducted on campus.The results show that the proposed dynamic clustering algorithm can effectively reduce the erroneous results,such as 70.60%of over-clustering and 49.76%of under-clustering,caused by the fixed neighborhoods of density-based spatial clustering of applications with density-based spatial clustering of applications with noise(DBSCAN),therefore,effectively enhancing the comprehansive clustering performance of the algorithm.

关 键 词：智能汽车 目标检测 激光雷达 点云聚类 KITTI数据集 基于密度的噪声应用空间聚类(DBSCAN) 

分 类 号：U471[机械工程—车辆工程]