堆叠散乱目标的6D位姿估计和无序分拣

针对目标散乱堆叠场景下的机器人分拣问题，建立一种从目标筛选、识别到6D位姿估计的无序分拣系统。利用局部凸性连接方法将Kinect V2相机采集的堆叠散乱目标点云数据分割成单独的点云子集，定义抓取分数从中筛选出最上层未被遮挡的目标作为待抓取目标，保证机器人分拣目标时能从上至下进行抓取；针对不同种类目标的分拣需求，基于匹配相似度函数对三维目标进行识别并定位抓取点；融合截断最小二乘-半定松弛算法和最近点迭代算法，建立目标6D位姿估计模型，保证目标点云和模型点云重合率低情况下的精确配准。在自采数据上进行目标6D位姿估计实验以及机器人无序分拣实验，结果表明：提出的6D位姿估计方法相较于流行的几种方法，可以更快速、精确地获取目标的6D位姿，均方根距离误差<3.3 mm,均方根角度误差<5.6°;视觉处理时间远小于机械臂运动的时间，在实际场景中实现了机器人实时抓取的全过程。

6D pose estimation and unordered picking of stacked cluttered objects

Aiming at the problem of robotic picking in the scenario of stacked cluttered objects, an unordered picking system from target screening, recognition to 6D pose estimation was established. The Locally Convex Connected Patches method was used to segment the stacked cluttered objects collected by Kinect V2 camera into separate subsets of point cloud, and the uppermost unshaded object was selected as the target to be captured by defining the capture fraction, so as to ensure that robot could grasp the object from top to bottom. According to the picking requirements of different kinds of objects, 3d targets are identified and grasping points are located based on matching similarity function. An object 6D pose estimation model is established by combining TEASER(Truncated least squares Estimation And SEmidefinite Relaxation) algorithm and ICP(Iterative Closest Point) algorithm to ensure accurate registration of target point cloud and model point cloud under low coincidence rate. Experiments of 6D pose estimation and robotic unordered picking are carried out on self-collected data. The results show that the proposed 6D pose estimation method can obtain the 6D pose of the target more quickly and accurately compared with several popular methods. The root mean square distance error is less than 3.3 mm and the root mean square angle error is less than 5.6°. The visual processing time is far less than the movement time of robot arm, and the whole process of robotic real-time grasping is accomplished in the actual scene.