基于点云-模型匹配的激光雷达目标识别

摘要

For the lidar automatic target recognition purpose, a method based on point clouds-model matching is proposed in this paper. The point clouds are projected to 3-views and binary images are created after binarization. The Sobel operator and Hough transform are applied to the binary images to extract the silhouette boundary and line parameters, then the bounding rectangles are extracted with the restriction of projected point clouds' silhouette information, thus the target's orientation and geometric feature are obtained. Taking the minimum of the mean Euclidean distance between point clouds and facets of the CAD model as the optimization target, utilizing the unit quaternion method for rigid transform calculation, the matching between point clouds and the CAD model is realized by using iteration, and the recognition process is finished by taking the normalized mean Euclidean distance as similarity metric. Point clouds of five kinds of ground armored vehicles in different lidar viewing angles are used to fulfill the recognition experiment. The statistical results show that the target's class recognition rate is 100% and the type recognition rate is above 915%. So this method performs well on recognition and has better application prospect.%为实现激光雷达自动目标识别,本文给出了一种基于点云模型匹配的方法.将点云进行三视投影,对投影点云进行二值化处理得到二值图像,采用Sobel算子和Hough变换提取点云轮廓边界及获得边界直线参数,然后以投影点云轮廓信息为约束提取包围矩形,完成目标姿态估计和几何特征提取.在此基础上以点云到CAD模型面元的欧氏距离最小为优化目标,采用单位四元数法计算点云与模型之间的刚体变换,通过迭代实现点云和候选目标CAD模型的匹配,并以归一化平均欧氏距离作为相似性度量完成目标识别.采用五种地面装甲目标在不同激光雷达视角下的点云进行目标识别实验,统计结果表明目标类别的正确识别率为100％,目标型号的正确识别率大于91％,因而本文方法具有较好的识别性能和较高的应用价值.