Research On Manipulator Grasping System Based On Point Cloud Pose Estimation Production Process

In recent years,with the development of intelligent robot technology,more and more robots have replaced the production activities originally realized by human.One of the more common tasks is to estimation the scattered industrial parts in the materiel section one by one and grasp them by a manipulator.The main target is to estimate the solve the 6D pose of the workpiece instance.Usually,it includes instance segmentation and pose estimation tasks,And whether the pose estimation of parts could be correctly is important to the grasping resultsThe traditional 6D pose estimation algorithm like based on RGB data algorithms need the object to have rich texture information,and it can’t deal with the information loss caused by occlusion.In the research field of point cloud,main of pose estimation methods often focus on large indoor or outdoor data by living objects,and interest in the directions of automatic driving,life application and so on.In the industrial scene,there is still widely use simple and traditional optimization algorithms.As the shape of the object becomes more complex and the task requirements become more strict,the accurate grasping of the manipulator is facing more difficult challenges.Based on the industrial scene,using point cloud data,combined with point pair features and deep learning,this paper studies the pose estimation and point cloud segmentation algorithm of the manipulator grasping system,and builds complete software and hardware system.The main research work and innovations are as follows:(1)In the stage of point cloud segmentation,aiming at the lack of imaging in the parallel projection direction of 3D camera,propose a two-dimensional and three-dimensional fusion based European clustering,and used a accurate k-means algorithm to the edge point compensation stage.The test result of self-built data sets shows that our method proposed in this paper has better performance than the traditional European clustering method,and is more suitable for point cloud segmentation in industrial environment.(2)The query structure based on point pair feature algorithm PPF is improved,used double-layer hash table to replace the original single-layer,according to the geometric characteristics of industrial parts,delete the rotation angle remove-repeat step in the extended neighborhood stage,and used the weighted voting method.Get better results than original algorithm on self-built dataset and modelnet40 dataset.(3)Proved the influence of initial rotation attitude to the final result in the rasterization stage of NDT pose optimization,and proposed an optimal rasterization method based on BFGS.In the off-line stage,the model is used to search the optimal rasterization pose,which achieves better results than random pose,and has obvious effect on some objects with specific geometric structure.(4)Based on PCRNet pose estimation network,we used Point Net++ to replace the original Point Net module,which improves the overall performance and enhances the robustness under the condition of missing occlusion.The iterative process of PCRNet is improved to an efficiently iterative process.