智能车辆3-D点云快速分割方法  

Fast segmentation of 3-D point clouds for intelligent vehicles

作　　者：王肖[1,2] 王建强[1] 李克强[1] 徐成[1] 李晓飞[1] 

机构地区：[1]清华大学汽车安全与节能国家重点实验室,北京100084 [2]军事交通学院汽车工程系,天津300161

出　　处：《清华大学学报（自然科学版）》2014年第11期1440-1446,共7页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(51175290);戴姆勒-清华联合项目

摘　　要：针对于智能车辆环境感知实时性要求,研究一种基于3-D全景式激光雷达的点云快速分割方法。采用带地面状态变化判断的分块直线拟合算法来去除背景地面数据,生成前景点云图像并利用机器视觉算法进行聚类分割。建立并利用坐标映射模型计算出定向包围体参数。在3组基于城市道路环境的测试集下,该方法同栅格法相比能够较好地降低过分割与欠分割错误率,其中车辆分割准确率约为95%,行人分割准确率约为85%,且耗时不受场景复杂度影响,能够较稳定控制在55ms/帧左右。This paper describes a fast segmentation method for 3-D point clouds targeted at real-time performance of intelligent vehicles. Efficiency is achieved in this approach by splitting the segmentation problem into two subproblems to remove the ground data and to seg- ment the foreground data using machine vision algorithms. A coordi- nate mapping model is then used to calculate the oriented bounding- box representing the target＇s shaper position and direction. The method＇s performance is evaluated on real data acquired in complex traffic environments. Test results show that the accuracy of the ve- hicle segmentation is 95% and that of pedestrians is 85%, with an average of 55 ms needed for each frame with the runtime unrelated to the traffic environment.

关 键 词：智能车辆 3-D LIDAR 点云分割 映射模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]