Object Recognition And Localization Based On Monocular Vision

Machine vision is always a hot topic, and varieties of applications based on that have been proposed. Object recognition and localization are the main problems during the study of machine vision, and the accuracy and complexity are important indicators of a computer vision system. Because of the advantages of simplicity and flexibility of monocular vision system, a monocular-based object recognition and localization system was proposed during the process of doing research on the basic theoretical concepts and the corresponding algorithms of it.Main research works include:firstly, the background, significance and application of machine vision are introduced, and the status of object recognition and localization is detailed, moreover, algorithms used during the process are briefly introduced and analyzed; Secondly, a perspective distortion correction method for objects’surface under monocular vision is presented. The properties of the vanishing line are used to obtain the coordinate transformation matrix. And the matrix indicates the relationship between image plane and object plane, according which, the process of image interpolation is realized. The results show its accuracy and efficiency. Thirdly, several algorithms of image feature extraction and description are briefly described, including their merits and demerits. In order to meet the real-time requirements, classify the extracted feature points of SURF (Speeded Up Robust Features) algorithm, then binding BBF search algorithm for matching feature points. Furthermore, the programming experimental results of the object recognition based on the improved SURF combined with BBF show that the algorithm proposed is fast and efficient in object recognizing. Fourthly, a couple kinds of coordinate systems during calibration and camera imaging models are depicted, and then, Zhang Zhengyou calibrating is detailed, with which the calibration is realized. Fifthly, knowledge on epipolar geometry and essential matrix are applied on processing the matching points after recognizing targets, and then the relationship between the current camera location and the previous one that used for building image library can be obtained from the essential matrix. Combining with camera’s internal parameters and prior knowledge, the location of the object relative to the camera can be obtained. The experimental results show that, the method we proposed can achieve the object location precisely, and the object localization was realized.The last but not the least, a conclusion of the paper is given, and a brief analysis of the follow-up study of monocular-based object recognition and localization system is added.