融合激光和视觉的局部人体近邻算法研究  

Research on Local Neighbor Algorithm Combining Laser and Vision

作　　者：范鹏鹏 陈万米[1] 李帅 

机构地区：[1]上海大学机电工程与自动化学院,上海200444

出　　处：《工业控制计算机》2021年第5期71-73,共3页Industrial Control Computer

摘　　要：环境感知技术一直以来都是无人驾驶邻域研究热点。由于激光雷达精确度较高的优点,通过对激光雷达的点云数据进行聚类划分的目标检测成为当今环境感知研究的热点技术之一。然而由于激光雷达数据的特征单一,导致其在对近邻人体检测时存在着误聚类的问题。以致于单靠激光雷达聚类进行全局聚类难以满足无人驾驶对车辆前方人体进行精确检测的需求。为了解决激光雷达存在的近邻区域人体误聚类的问题,提出了一种融合激光和视觉信息的近邻聚类算法。此算法主要是首先以自适应欧氏聚类算法对激光点云图像进行全局聚类,并通过立体视觉进行图像匹配和目标检测获取车辆前方人体三维坐标,从中提取人体之间满足近邻标准的区域和人体目标个数,并与激光雷达中原有聚类进行近邻区域匹配,最后以k均值聚类对激光雷达点云中的局部邻近区域进行点云划分,以此得到更为准确的车辆前方的聚类信息。最后通过实验证明了此算法可有效地解决激光雷达在环境感知方面存在的近邻区域误聚类的问题,提升了激光雷达环境感知方面的鲁棒性和准确性。In order to solve the problem of human body misclustering in the near neighborhood area existing in lidar,this paper proposes a near neighborhood clustering algorithm which integrates laser and visual information.This algorithm is mainly use the adaptive european-style clustering algorithm to global clustering of laser point cloud image,and through the stereo image matching and target detection for the vehicle in front of the human body three-dimensional coordinates,extract meet neighbor standard zone between human and the human body target number,and compared with original clustering in the laser radar neighbor region matching,finally with k-means clustering to local adjacent area of lidar point cloud point cloud classification,in order to get more accurate information of the vehicle in front of clustering.Finally,the experiment proves that this algorithm can effectively solve the problem of near-neighborhood region misclustering in the environment perception of lidar,and improve the robustness and accuracy of the environment perception of lidar.

关 键 词：全局聚类 目标检测 近邻区域 局部聚类 

分 类 号：U463.6[机械工程—车辆工程] TN957.52[交通运输工程—载运工具运用工程]