Research On Key Technologies Of Vision Positioning For Nondestructive Testing Robot

One approach to non-destructive testing is to combine vision technology with robotics.Stereo vision technology can obtain the depth information of the scene target,and realize the recognition and positioning of the detection target.In order to solve the problem that the poor matching accuracy affects the visual positioning effect in the actual detection process,this paper has carried out research on the key technology of visual positioning of NDT robot.The specific research contents are as follows:(1)According to the research content of this topic,firstly,the research status of stereo matching and point cloud matching are introduced and analyzed in detail.On this basis,the frame design and the whole processing process of robot vision positioning system are determined,and the main hardware and software platform of this system are described in detail.(2)The camera imaging model and principle are introduced.Based on this,the camera calibration experiment is carried out,and the conversion relationship between the three-dimensional coordinates of a certain point of the scene target and its corresponding point on the image is established According to the concept and principle of hand-eye calibration,the calibration experiment of eye-out-of-hand system is carried out to make the robot arm accurately locate the points on the image.(3)The acquisition method and pretreatment of scene target point cloud are described in detail Firstly,the process of reconstructing sparse point cloud by motion reconstruction SFM algorithm is introduced.Sift algorithm is used to extract and match feature key points.For mismatched feature points,RANSAC algorithm is used to optimize and improve matching accuracy.Then MVS algorithm is used to add matching images and re-delete point sets to generate patch model and reconstruct dense point cloud;Finally,the scene object is reconstructed to restore its surface detail texture.(4)A point cloud matching algorithm model based on deep learning is designed In order to reduce the influence of non-overlapping points on the matching accuracy,an overlapping mask is used to segment the overlapping region of the target so that the non-overlapping region does not participate in the pose matching process Firstly,Edge Conv is used to extract features.Secondly,attention mechanism is added to enhance the feature interaction between point clouds.Then,point features are fused with the global features of view point cloud and scene target point cloud.Finally,pose regression is carried out to output pose.(5)Finally,using the existing experimental platform to verify the accuracy of pose estimation Experimental results and error analysis show that the position estimation error is below 6.5 mm and the attitude estimation error is below 4.6°.