基于视点特征直方图的激光点云模型的位姿估计  

A posture estimation method based on viewpoint feature histogram for laser point cloud model

作　　者：张彪[1] 曹其新[1] 焦瑶[1] 

机构地区：[1]上海交通大学机器人研究所,上海200240

出　　处：《光电子．激光》2013年第7期1357-1362,共6页Journal of Optoelectronics·Laser

基　　金：国家高技术研究发展计划"863"(2012AA100906);教育部重大项目培育(708035);上海市教委创新(12ZZ014)资助项目

摘　　要：提出一种基于视点特征直方图的点云模型位姿估计算法。首先在目标物体周围采集三维点云,拼接后获得物体的完整点云模型;然后对点云模型计算其视点特征直方图,构建特征数据库;对待估计点云同样计算其特征直方图,使用KNN算法在数据库中搜索与之最接近的位姿作为初始位姿估计值;最后使用迭代最近点(ICP)算法将待估计点云精确配准到模型点云,从而获得坐标系之间的相对位姿。实验表明,这种方法对于物体位姿识别有很强的稳健性,能很好实现目标物体的三维位姿计算。Robots have the demand to interact with different objects within their expected visual range, thus precise pose estimation of target objects is in urgent needs. In this article, a novel pose estimation method is proposed based on viewpoint feature histogram and ICP algorithm using 3D point cloud data gathered by laser range finder （LRF）. Firstly,point cloud data is collected around the target object,then we register different point clouds together and get the complete point cloud model for the target object. Next,we calculate viewpoint feature histogram for every point cloud model and build a database filled with feature histograms from different views. Once we get a point cloud from a new position, we can search the most appropriate candidate in the database using KNN algorithm, and use it as initial value of posture matri~ Finally, ICP algorithm is utilized to minimize registration error and get the precisely esti- mated posture matrix,and the whole process stops when the result is precise enough or the number of it- erations exceeds the limit. We test this method with both simulations and experiments, and the results that this method is of strong robustness when estimating posture of the target objects.

关 键 词：信号处理 位姿估计 视点特征直方图 激光测距仪(LRF) 三维点云 

分 类 号：TN911.74[电子电信—通信与信息系统]