Binocular Image Matching Algorithm Based On Color Anisotropic SIFT Features

Binocular vision system obtain three-dimensional(3D)information of objects based on the parallax principle,which is a major research hotspot in the field of computer vision.The matching algorithm of binocular image is one of the important steps in the implementation of binocular vision system.The feature-based matching algorithm has two important steps: one step is feature extraction.The distinctions between feature vectors and the robustness of feature vectors have a great impact on subsequent matching performance;another step is feature matching.Excellent matching measure is very important.In this thesis,proposing a bincular image matching algorithm that based on color anisotropic SIFT features.The main work is as follows:(1)SIFT features are a relatively stable point features,and their descriptors have good invariance to scale,illumination and affine,etc.SIFT features are based on grayscale images,without using color information of image.In addition,SIFT features use Do G operators to construct pyramids,which do not make use of the anisotropy of point features.This thesis proposes a color anisotropic SIFT features extraction algorithm.Firstly,the color fusion model is applied to the pyramid,and the Do G operator is replaced by the anisotropic Lo G operator,making full use of the color information of the image and the anisotropy of the point features.Secondly,in order to obtain more accurate gradient and direction information while constructing descriptors,this thesis uses the anisotropic Gaussian directional derivative filter instead of the Sobel operator.With the matching experiments on binocular images which have complex background and repeated textures in complex scenes,it is shown that using color anisotropic SIFT features still can achieve higher matching accuracy and robustness to noise.(2)For the matching algorithm of SIFT features,the matching strategy usually adopts the nearest neighbor distance ratio algorithm which is based on Euclidean distance.However,the correct matching points appear in the k neighbor with a high probability.There may be a large number of mismatch points in this algorithm.This thesis proposes a matching algorithm which uses color anisotropic SIFT features as matching primitives.Firstly,the k nearest neighbor features distance ratio and cosine similarity are used to replace the nearest neighbor distance ratio as matching measures to reduce mismatch points.Secondly, in order to remove mismatch points and outliers,we adopt spline curve loss instead of 0-1 loss in RANSAC algorithm to reduce the influence of threshold.Meanwhile,Sampson is used as error function.Above all,it can obtain matching relationship which remove mismatch points and outliers.Finally,the iterative nearest neighbor algorithm has higher requirements for the initial matching relationship of the point set.The matching relationship mentioned above with higher accuracy is used as the initial matching relationship of the algorithm,it can reduces the iteration times and matching errors.With a large number of experiments on binocular color images,it is verified that the matching algorithm proposed in this thesis has higher matching accuracy and lower matching error.