Design And Implementation Of Object Recognition And Pose Estimation Based On Binocular Vision

Object recognition?location?tracking and pose estimation are important front-end technologies for intelligent grasping of manipulator.it's recognition rate,location accuracy,tracking reliability and position accuracy directly affect the grasping of the back end manipulator.With the maturity of computer vision technology,binocular vision recognition technology has been applied in a wide range of fields.In this paper,ZED binocular vision camera is used to identify target objects,position attitude estimation,point cloud matching and other related operations.The paper firstly explains the research background and significance of the topic,reviews and explicates the relevant domestic and foreign literatures,briefly introduces the related main functions involved in the project,and explains the minimum hardware configuration used by the ZED camera.Secondly,the basic pinhole camera model is introduced,the lens distortion and the homography matrix are described,and a single target experiment is performed to obtain the single target parameters.The principle of stereo imaging is explained.The principles of triangulation,the principle of solving the eigenmatrix and the fundamental matrix,the principle of stereo calibration,the principle of stereo correction and the principle of stereo matching are respectively expounded.The binocular calibration experiment was carried out,and the binocular calibration parameters were obtained,which provided a theoretical basis for subsequent object identification,matching and tracking.Then the paper carries out the scheme design,implementation and experiment based on the contour identification of planar low texture object,and analyzes the reliability of each specific experimental environment.The characteristic points recognition based on planar complex texture object program design and experiment,discusses the real-time performance of the scheme,and uses the CamShift algorithm for general background environment is used to identify the track of the objects at low speed.Finally,in the 3D desktop environment,the ZED SDK is used to obtain the depth image and point cloud.It is used to filter,segment and cluster the point cloud,and construct the general position measurement experiment platform to obtain the target recognition point cloud training set.Based on VFH descriptors,a supervised machine learning recognition algorithm is used to train,test and analyze point cloud training set,and the reliability of the experiment is described.Finally,the thesis summarizes the overall work and explains the main personal work.In addition,the target of FANUC LR Mate 200 iD 6-dof manipulator was studied,and further work was prospected.