基于kinect的物体抓取场景认知

针对RoboCup竞赛家庭组比赛对物体抓取的要求,研究了物体抓取场景认知问题。设计了一种基于kinect的物体抓取场景认知系统。先将kinect传感器得到的深度图像转换为3维(3D)点云图,然后计算每个3D点所在曲面的局部法向量,再根据法向量和距离特征分割提取出水平桌面;采用3D点与水平桌面的位置关系分离出潜在的抓取物体目标点,选择随机抽样一致(RANSAC,Random Sample Consensus)算法完成圆柱形抓取物体的定位。使用实验室采集的场景深度图对认知系统进行测试。结果表明,设计的系统可以可靠提取水平桌面和桌面上的圆柱形物体,可以达到物体抓取比赛的要求。

Scene cognition for object grasping based on kinect

Scene cognition for object grasping is studied,according to the desired ability of object grasping in @ Home Leagues of RoboCup.A scene cognition system for object grasping is designed by using a kinect.First,the depth image from a kinect sensor is transformed into 3D point clouds.Second,the local surface normal vectors are computed for every 3D point.Then the horizontal table planes are segmented out with normal vectors and distance features.The potential target points of graspable object are detected by the position relation of 3D point and the horizontal table plane.Random Sample Consensus（RANSAC） algorithm is selected to perform the localization of graspable cylindrical object.The proposed cognition system is tested using the depth images of scene captured from our Robot Lab.The results show that horizontal table planes and the cylindrical objects on the table can be reliably extracted,and it can meet the demand for object grasping of @ Home Leagues of RoboCup.