Research On Heteroplane Target Parallel Pose Acquisition Based On Binocular Vision And CAD Model’

Under the multi-variety,small batch,customized,and personalized production mode,the teaching-type industrial robot can no longer adapt to it,and needs an intelligent robot with visual functions.In order to effectively complete the overall motion planning of the intelligent robot and the operation planning of the local to-be-processed/assembled parts with no\weak-textured workpieces in the industrial site,it is necessary to obtain the postures of the different parts of the workpiece and the local to-be-processed/assembled parts at the same time.Vision-based posture acquisition of workpieces can generally be achieved through two methods:monocular and binocular.Monocular mostly adopts CAD template method,which is to measure the overall position and pose,which cannot meet the position of the different plane parts to be processed/assembled overall and local Pose is acquired in parallel;binoculars depend on the feature information in the camera’s angle of view,and cannot directly acquire the pose information of blind spots in the visual field and multiple parts on different surfaces.This paper takes the non-weakly textured workpiece in the field of mechanical manufacturing as the research object,based on the method of combining binocular vision and 3D CAD model information,researches on the problem of parallel position and pose acquisition of multi-part targets on different surfaces of the workpiece,and proposes a point cloud based Sampling improved ICP registration algorithm to obtain a transformation model between the CAD model of the workpiece and the visual coordinate system of the robot on the job site,and convert the posture and posture of the part to be processed/assembled in the CAD model through the transformation model.The research content is mainly aimed at the three initial elements required for the registration of the transformation model,namely the initial relationship,the solution of the visual and model point clouds,and the expression of the pose information.The specific research results and innovative contributions are as follows:1.Solve the initial relationship of registration based on the two-dimensional contour of the visual image and CAD model.To solve the problem of too few visual point clouds and the inability to obtain the initial relationship required for registration through the point cloud centroid,the two-dimensional contour similarity matching of images was used to obtain the initial relationship required for registration.First,filter,highlight suppression,and edge extraction on the workpiece image to effectively suppress the effect of noise and surface reflection on edge extraction;then,obtain a two-dimensional projection by setting a virtual camera in the CAD model space of the workpiece,and extract straight edges;through similar Sexual matching obtains two-dimensional corresponding pairs of straight lines,and finally obtains the initial relationship with the three-dimensional information of the straight lines.2.Solve the transformation model based on binocular vision and 3D CAD model.First,for the homography problem of binocular visual contour matching,the affine model is used to solve the problem of homography matching of weakly textured workpieces,and the epipolar constraint method is proposed to solve the problem of homography matching of non-textured workpiece contours.The epipolar constraint method is less accurate than the existing methods,but it can meet the matching requirements of non-textured workpieces;then,read the three-dimensional point cloud of the CAD model in STL format,and perform deduplication and coarse sampling;finally,for the transformation model For the problem of different quasi-midpoint cloud spacing,an improved ICP registration algorithm based on point cloud resampling is proposed.This algorithm uses the initial relationship and iteration direction as a priori conditions for rough registration,and the corresponding model point cloud is obtained according to the contour points.Re-sampling the minimum pitch of clouds can effectively solve the problem of different pitches of the two parts of the point cloud in registration.3.CAD space posture information conversion based on transformation model.Aiming at the problem of expression and conversion of complex pose information in CAD models,an index model based on quadtree is proposed.This method uses the type of workpiece to be assembled,the type of assembly part,the assembly sequence of the assembly part and the assembly position pose information to establish a four-dimensional The digital model uses the three-dimensional information of the model to establish the quadtree structure as the index number,and the fourth-dimensional information as the pose index file in the model index library.The establishment of the index model makes the posture information and assembly information clearer,more hierarchical,and more efficient in assembly.4.Experiments on accuracy verification of parallel pose acquisition of heteroplane targets based on binocular vision and CAD model.In order to avoid the operation error caused by directly measuring the absolute pose of the workpiece target in the robot coordinate system,the relative measurement method is used to translate and rotate the experimental workpiece,and the relative amount between the two poses is used to calculate the error.Experiments show that the maximum errors of obtaining poses are 1.209mm and 0.018o.Compared with existing methods,it is proved that the pose method proposed in this paper can simultaneously obtain different positions and positions of different parts and can meet the general robot grabbing tasks,most Precision requirements for assembly and palletizing tasks.