Study On Object Tracking And Localization Algorithm Based On Robot Vision

Machine vision is exploration and simulation of human vision by computer technology. With the development of computer technology, machine vision has been applied to manufacturing and life widely. Machine vision includes target detection, target tracking, target localization, etc., it is a technology which based on image information from cameras. In this thesis, those key issues are researched and implemented.Initially, to the deficiencies of moving target detecting algorithms, such as, be vulnerable to the lightness and the object with similar color, etc., this thesis proposes a kind of moving target detecting algorithm based on the space edge direction histogram(SEOH). The algorithm improves those deficiencies. Further, combining with the with the Kalman filter, the real-time of the algorithm is improved.Subsequently, to the deficiencies of Camshift algorithm, such as, sensitive to similar color, target occlusion, target losing, etc. This thesis proposes an improved Camshift algorithm based on the space edge direction histogram ( SEOH ) . The algorithm improves the capability of anti-interference to color, obtains better effectiveness by considering block matching in the case of target occlusion. Moreover, by predicting and updating the position of the target using Kalman filter, the real-time of the algorithm is improved.Consequently, on the base of summarizing and analyzing of the existing camera calibration methods, this thesis researches the issue of the calibration of different cameras. The calibration processes are as follows: first, building binding equations according the corresponding relation among calibration boards; then, on the basis of obtaining the intrinsic parameters by these equations, calibrating the relative position of different cameras. Experiments proves the effectiveness and practicality of the algorithm.Finally, this thesis proposes an algorithm which combined pixel value and image features. A kind of rough matching algorithm based on pixel value is proposed to initializing the matching area. Then, by considering the similarity of key points in different scale spaces, the SIFT algorithm is improved through rejecting the unstable feature points and the weak edge points. Experiments show that both the accuracy and the real-time of images matching are improved. Moreover, using the calibration results and the matched feature points, the 3D information of the target are recovered by binocular disparity principle. Localization results show the accuracy and practicality of the binocular vision localization system. This thesis researches some key issues in machine vision, such as moving target detecting, moving target localization, camera calibration and image matching, etc. More, in the circumstance of lab , the algorithms are implemented.