一种非结构环境下目标识别和3D位姿估计方法  

A novel method of target recognition and 3D pose estimation in unstructured environment

作　　者：任秉银[1] 魏坤[1] 代勇[1] REN Bingyin;WEI Kun;DAI Yong(School of Mechatronics Engineering, Harbin Institute of Technology, Harbin 150001, China)

机构地区：[1]哈尔滨工业大学机电工程学院,哈尔滨150001

出　　处：《哈尔滨工业大学学报》2019年第1期38-44,共7页Journal of Harbin Institute of Technology

摘　　要：为提高非结构环境下目标识别准确率和位姿估计精度,提出一种利用Kinect V2 RGB-D传感器并基于目标的CAD模型进行不同类型目标自动识别和3D位姿估计的新方法.利用虚拟相机获取目标CAD模型的深度图像,并将目标的模型转化为点云图,采用体素栅格滤波减少场景点云中的点数;利用点对特征描述子(PPF)作为CAD模型的全局描述子,并将相似的PPF划分成一组放进一个hash表,用于识别和定位目标,所有目标的hash表组成了3D模型数据库;利用基于投票策略的方法对不同类型目标进行检测识别和3D位姿估计,并采用位姿聚类的方法和ICP配准进行位姿修正,再通过奇异值滤波剔除误匹配位姿,从而提高位姿估计精度.在虚拟机器人实验平台仿真环境中分析了3种管接头的识别率和位姿估计误差,结果表明:3种管接头平均识别率96%,位置误差<4 mm,姿态误差<2°,能够满足机械臂抓取要求.将提出的方法与两种主流位姿估计方法进行了对比实验,结果表明,提出的方法无论是识别率还是F1分数都要优于其他两种方法.To improve target recognition accuracy and pose estimation precision in unstructured environment,a novel method for automatic recognition and 3D pose estimation of different kinds of targets is presented based on the object CAD model and Kinect V2 RGB-D sensor.The depth image of object CAD model is obtained by a virtual camera,and it is converted into one point cloud.A voxel grid filter is utilized to reduce the number of points for the point cloud of the scene,and the point pair feature(PPF)is used as the global descriptor of CAD model.The similar PPFs are classified as the same group and put into one hash table to recognize and locate the targets.The hash tables of all targets compose the 3D model database.A voting scheme is adopted for different kinds of objects recognition and 3D pose estimation.Pose cluster and ICP registration are employed to refine 3D pose.The accuracy of 3D pose estimation is improved by filtering the mismatching poses.The recognition rate and the pose estimation error of three kinds of pipe joints are analyzed in the environment of virtual robot experimentation platform(V-REP).The simulation results show that the average recognition rate of three kinds of pipe joints is 96%.The position error is less than 4 mm and the orientation error is no more than 2°,which meet the requirement of manipulator grasping.The experiments of comparison with other mainstream methods are performed.The experimental results show that the recognition rate and the F1 score of the proposed method are superior to those of two other methods.

关 键 词：目标识别 3D位姿估计 非结构环境 CAD模型 位姿聚类 ICP配准 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]