基于改进DBSCAN算法的激光雷达水面目标检测

Detection of Water Surface Targets by Lidar Based on Improved DBSCAN Algorithm

激光雷达点云是一种目标近距离密度大、远距离密度小的不均匀点云.DBSCAN聚类算法有邻域值ε和密度阈值Minpts两个重要参数,参数的选取比较困难.传统的DBSCAN算法,采用固定的邻域值ε和密度阈值Minpts难以对全部数据实现良好聚类,对密度不均匀的激光雷达点云数据集会出现近距离的目标欠分割、远距离的目标漏检的情况.针对上述两个问题,提出了一种自适应参数的改进DBSCAN算法,对每一个不同的点云数据采用独立的ε进行聚类.确定激光雷达相邻两根扫描线距离,再乘以相应的点云距离作为每一个点云的邻域值ε.计算点云每一个点的包含在ε邻域内的点数,统计点数的期望作为Minpts.通过统计近距离出现过分割或视为噪声的目标点云,计算其邻域值ε的最大值获得邻域值下限,防止近距离点云邻域值过小导致过分割的情况.仿真结果表明:改进的DBSCAN算法,既能够区分出近距离的障碍物,也可以对远距离的目标进行聚类.

Lidar point cloud is a kind of uneven point cloud with high target density in short distance and low target density in long distance.DBSCAN clustering algorithm has two important parameters,neighborhood valueεand density threshold Minpts,which are difficult to select.Traditional DBSCAN algorithm,with fixed neighborhood valueεand density threshold Minpts,is difficult to achieve good clustering for all data.For the laser radar point cloud data set with uneven density,the short-range target is under-segmented,and the long-range target is missed.To solve the above two problems,an improved DBSCAN algorithm with adaptive parameters was proposed,which used independent neighborhood values to cluster each different point cloud data.The distance between two adjacent scanning lines of lidar was determined,and then multiplied by the corresponding point cloud distance as the neighborhood value of each point cloud.The number of points contained in the neighborhood was calculated,and the expectation of statistical points was taken as Minpts.By counting the target point clouds that are over-segmented or regarded as noise at close range,the maximum value of its neighborhood value was calculated to obtain the lower limit of neighborhood value,so as to prevent over-segmentation caused by too small neighborhood value of close range point clouds.The simulation results show that the improved DBSCAN algorithm can not only distinguish the obstacles in the short distance,but also cluster the targets in the long distance.