| In order to patrol the substation by using an autonomous robot,various advanced technologies are applied to power system to try to develop out the substation intelligent patrol robot system with the ability to entirely percept environment and process intelligent information.How rapidly and exactly to percept driving environment is always a hotspot in the research of artificial intelligent.However,the wide application of real-time vision navigation and position as well as 3D reconstruction is confronted with many difficulties because of the complex of images.Currently,the stereo vision technology directly simulates the way that animal handles the scenery,aimed to make a computer have the ability to recognize 3D environment information through 2D images.By taking substation intelligent patrol robot binocular stereo vision system as research object,some methods and problems of acquiring depth information using binocular vision are researched,calibration of binocular imaging model parameters,estimation of binocular disparity and preceding environment perception technology of robot based on vision are studied in this dissertation.The primary work include:Firstly,the paper introduces the linear perspective projection model of single camera and the nonlinear perspective projection model considering the radial and tangential deviations,Based on binocular stereo vision Parallax,the theoretical basis for obtaining environmental depth information by triangulation method is analyzed,and the transformation relation between the optical imaging and the different coordinates of the cameras is clarified.Meanwhile,a general binocular stereo vision System is constructed,and the coordinate transformation of the left and right camera is deduced.Finally,the geometrical model and the polar Line correction technique of binocular stereo vision are fully explored.In the thesis,the camera calibration process is analyzed in detail,and the influence of lens distortion is discussed,the calibration result is corrected.The author designed and tested the calibration procedure of the camera model based on OpenCV to calibrate camera parameters.Through the test of four different types of checkerboard images,the projection error,calibration accuracy and polar line correction analysis of the camera calibration result are verified,and the designed program can calibrate binocular stereo vision camera accurately.The method of disparity estimation is studied in depth in this thesis,and the combination of a method of classification-hypothesis-verification and Levenberg-Marquardt(LM)algorithm is to extract candidate disparity points after analyzing characters of disparity points and gray characteristics of their adjacent areas,and then the set of disparity points were optimized by using Dynamic programming algorithm.Aiming at the occlusion problem appearing during the generation of stereo parallax graph and the problem of the stripe defect in stereo matching only using dynamic programming algorithm,based on the basic theory of stereo vision matching,a stereo matching algorithm based on piecewise parabolic-fitting is proposed to fit a optimal parallax function and to improve the accuracy of disparity estimation.Comparing with the stereo matching performance of several typical algorithms,the validity and timeliness of the proposed algorithm are verified by experimental results.In order to further improve the stereo matching of binocular stereo vision system,the paper deeply analyzed the basic principle of Maximum Flow-Minimum Cut Algorithm,which is to obtain the global optimization and introduced which to the stereo image matching optimization.Network is established using disparity labels,the algorithm will be simplified through replacing the smoothing item with the difference distance of pixels in order to realize the optimal solution of the energy function based on the couple of epipolars,thus binocular stereo matching image achieved success with Maximum Flow-Minimum Cut Algorithm.By using the couples of standard images in the Middlebury website tested in,the results show that Maximum Flow-Minimum Cut Algorithm has better matching performance compared with stereo matching algorithm based on parabolic-fitting.Moreover,some correlative algorithms of image preprocessing are analyzed in the thesis in the aspect of improving the real-time and robustness of the system.For instance,median filtering method is used on removing the noise,Sobel operator of different nucleus is used to enhance edges of lanes and occlusions,the method of maximal variance between-class is used on extraction of features of lanes,as well as the adaptive double-threshold method is for recognizing the bottom edge features of occlusions.By preprocessing the image,the accuracy of the target recognition is improved and the calculation will become much easier,then the robustness of the algorithm is ensured while improving the real-time performance of the algorithm.At last,Hough Transform is used in the lane detection in this thesis,after the 3D lanes are reconstructed using the estimated disparities of extracted 2D lanes of left and right image,the relative position between obstacles and lanes are determined.And then statistic prediction method and Kalman filter technology are used in lane tracking.The feature points of obstacle are reconstructed using their estimated disparities,and then the distance between obstacle and host robot are determined. |
