Research On Identification And Location Of Textureless Industrial Workpieces Based On Robot Eye-hand System

Workpiece assembly is an indispensable component in the field of intelligent manufacturing.Nowadays,the production model of small batches and diversification places higher demands on the speed,accuracy and flexibility of assembly work.Compared with the traditional method,the intelligent assembly based on collaborative robots ensures the product quality while maximizing the production flexibility and efficiency.Research on intelligent assembly technology is mainly focused on identification and location of industrial workpieces.In this paper,we take the textureless industrial workpieces with planar features as the research target,and take the identification of workpieces under complex background as well as the high-performance pose estimation as the key points.The research idea is proposed that transforms the recognition problem into two steps,which are planes extraction and primitives reorganization.The main research work of this paper is as follows:(1)Due to the limitation of depth-based plane segmentation,an optimized algorithm is proposed which combines the line features in the color map.What's more,the multi-view eye-hand system is applied to achieve high-performance planes extraction.Firstly,the low-cost RGB-D camera is used to obtain the images from the scene,and the pre-processing of images is applied to reduce the noise interference which is caused by highlight and depth noise.Secondly,the depth map segmentation algorithm is combined with Canny features by optimizing the hierarchical clustering process of the image nodes to a fast clustering process based on the gradient boundary,thereby reducing the running time of the plane segmentation algorithm and improving the accuracy of segmentation.Thirdly,with the advantage of eye-hand system,the plane primitives obtained from different perspectives are merged so that to improve the problem of occlusion or incompleteness in the stacking scene.(2)In order to solve the problem of plane features distortion,a data-driven workpiece identification algorithm is proposed.By training the primitives relationship characteristics,the primitives in complex background are reorganized so that each instance of workpieces is obtained.Firstly,the Gazebo simulation software is used to generate the dataset of virtual stacking scene,and the construction method of the primitive features is studied.Secondly,features selection and preprocessing methods are used to reduce the dimension of the features,and the processed data are trained respectively at Random Forest,K Nearest Neighbor and Support Vector Machine.The cross-validation experiments are tested for each classifier and the result shows that the Random Forest classifier has the best performance.Thirdly,a fast identification algorithm for stacking workpieces is proposed to output each instance of workpieces.(3)In order to achieve high performance of pose estimation in the stacking scene,a coarse-to-fine strategy based on global search method is proposed,so as to improve the running speed of the pose estimation algorithm on the premise of ensuring accuracy.In the coarse stage,the Super Key-4PCS registration algorithm which combines Harris features is proposed.This algorithm achieves fast registration on key-point sets with the intelligent indexing scheme and generate initial pose of the workpiece.In the refinement stage,the Iterative Closest Point method is used to obtain the precise pose of each workpiece on the basis of the initial pose.Finally,the proposed coarse-to-fine strategy is quantitatively evaluated on the Linemod dataset,and the comparison experiments are taken between Super Key-4PCS and some other excellent algorithms.(4)In order to combine planes extraction,workpiece identification,pose estimation and robot control for intelligent grabbing task,comprehensive experiment of stacking workpieces is designed on the basis of ROS framework.Firstly,the hardware platform is built including the selection of the depth camera and robot arm,as well as the design of the end effector.Secondly,the architecture of the pose measurement system is designed by using ROS to realize the interconnection of the three major modules.Finally,comprehensive experiments are carried out in virtual and real scenes,and we also use the eye-hand matrix to convert the poses of workpieces to the robot coordinate system to achieve grabbing experiments.