Research On Stereo Target Recognition And Pose Estimation Based On Template Matching

With the strategic direction of "Made in China 2025" proposed by China,robots on traditional assembly lines have a low fault tolerance rate and are difficult to content the requirements of flexible production in complex environments.Robots based on visual technology assisted are widely used in automated assembly,sorting and other fields.However,in actual production,it is difficult for robots to recognize target object and estimate pose under the condition of complex environment and mutual occlusion,which poses certain challenges to the current 3D object detection and recognition and pose estimate algorithms.In this paper,the color image and depth image collected by Kinect camera are used as information sources.Research on stereo target detection and pose estimation algorithms in complex environments,and the robotic arm is used to capture the target objects.The work of this paper mainly includes the following aspects:(1)In the composed robot system,the Kinect V2 vision sensor was first calibrated with Matlab2014 software,the internal parameter matrix of the color camera and depth camera was obtained,and the alignment experiments of the collected color image and depth image were performed by using Opencv.The eye-to-hand installation method is used to install the vision sensor at a fixed position in the scene,and perform Eye-to-hand calibration of the robot system.The Tsai two-step method is used to solve the matrix equations,finally the camera and robot coordinate system's transformation matrix can be obtained.(2)A template matching algorithm based on multiple features is studied.Aiming at the shortcomings of the Linemod algorithm for foreground occlusion and repeated detection,nonmaximum suppression and adaptive threshold selection methods are proposed.To extract the optimized real object edge of the target object,the lag threshold algorithm is used to supplement the edge information,which makes the edges more complete.The virtual camera in Open GL open image library is used to collect template images in various poses of the target object,and save the pose information to create a template library,eliminating the complicated process of manually collecting templates;the similarity estimation function is redefined according to the feature information,and the template with the highest score is taken as the matching result.Finally,the effectiveness of the improved algorithm is verified through comparison experiments.(3)The ICP precise registration algorithm for point cloud registration is studied.Based on the classic ICP iterative closest point algorithm,improvements are made to improve the accuracy and speed of point cloud registration.In the point cloud preprocessing stage,the RANSAC algorithm is used to fit the bottom equation of the box body,and the pass-through filter is used to remove redundant points,and the radius filter is used to remove outliers;the point cloud data was down-sampled using the voxel grid method;the topological structure of the point cloud data was achieved through KD-Tree.A quick search of the nearest point will obtain the exact target point cloud pose.