3D Reconstruction Of Research Based On Binocular Vision

Machine vision is an interdispline involving many subjects such as artificial intelligence,computer science, image processing and pattern recognition and so on. By simulating thephysiological structure of human vision, binocular stereo vision is featured by its excellencessuch as small size, low cost, low power waste and high efficiency and so on. Therefore it isconsidered to have a broad prospect in many areas such as the 3D measurement of spaceobjects. It has been the focus of current research and widely applied in various fields ofnational economic development. This paper introduces the research content of binocularstereo vision, application field and reviews the literature home and abroad. The goal of thispaper is to reconstruct 3D scene information by taking two 2D images through slight paralleltranslation of one camera. It can be divided into four parts: the camera calibration, featureextraction, corresponding points matching and 3D reconstruction.The thesis studies and analyzes the algorithm and method in every part. The main content ofeach part is described as following.The thesis firstly introduces the transformation relationship of four reference frames from theangle of ideal imaging model. It discusses the three distortion camera models of the nonlineardistortion. It describes several traditional calibration approaches of cameras including DLTcalibration method, calibration method of cross ratio invariance, Zhang’s calibration methodand Tsai two-step method. The thesis completes the calibration test of non-linear model ofradial distortion by using matlab toolbox with Zhang’s method.Then, the thesis introduces four representative operators of edge characteristics in details(sobel operator, prewitt operator, roberts operator, canny operator) and describes two angularextraction algorithm: Susan algorithm and Harris algorithm. The effects of four edgecharacteristic operators (sobel, prewitt, roberts and canny) are extracted on the basis ofexperimental verification and the feature comparison in designated“building blocks”picture.By setting different value of the threshold parameter of Harris algorithm, it is verified that theHarris algorithm is under the effects of the variance of threshold parameter. Then the thesisrefines the sift algorithm after its description by supplementing it with epipolar linerectification after extracting feature points. The experimental results show that the larger thefocus, the bigger the emendation degree.Then, the paper describes the principle and the classification of the three-dimensional featurematching algorithm. It also describes the related several measure functions and fiveconstraint conditions. The thesis introduces several matching algorithms including algorithmbased on the matching of the snake, algorithm based on the matching of the sliding window,algorithm based on the matching of the interest points. Basing on above discussion, this paperrealizes the matching of feature points by using an improved sift feature extraction algorithmand the processing of original image.The final step, this paper completes the whole process of binocular stereo imagereconstruction through interpolation and fitting.Compared with Harris angular point extraction algorithm and line feature extractionalgorithm, the SIFT feature extraction algorithm has the advantage of keeping invariance to rotation, scale, the change of brightness. To some extent, it has a relatively stable ability offeature matching which is invariant to the changes of affine and perspective. The experimentreaches the expected results.