Research On Feature Matching Algorithm For Uncorrected Fish-eye Image

Since fisheye lens has characteristics of large FOV,short focal length,etc,it has large advantages to build stero vision system.However,because of the large distortion of fisheye images caused by the optical characteristics of fisheye lens,common feature matching algorithm used for perspective projection images can not apply to the vision system built by fisheye lens.What's more,the conventional feature matching approach for fisheye images needs to converts the fisheye images to perspective projection images through distortion correct,in which the approach not only lose lots of image information,but also introduce imprecise variables.So it is need to explor an effective method for uncorrected fisheye images.This paper will respectively research the only matching point of interesting point and the feature matching method of whole image for uncorrected fisheye image pairs.The main works include:(1)This paper studied the feature matching method for interesting points between two fisheye images.To achieve the purpose of feature matching,a combined feature descriptor(CFD)is built and the corresponding matching strategy is designed.This descriptor combines with color histogram that contains color feature,local binary pattern(LBP)which contains texture information and weber local descriptor that contains direction information,so as to realize the description of interest point.On the matching strategy,first of all,interesting point A and a region DR include it in another image is selected;building a description region R(A)for A and partitioning DR with the size of R(A)at the same time;then to obtain the one dimensional feature descriptions these regions are described by CFD;finally,the similarity of histograms is used to find the only matching point for A.Experiments show that for the interest point located in the image edge with big distortion,the method can find the matching point effectively.(2)Yet for the whole image,this paper combines the scale invariant feature transform(SIFT)and center symmetry local binary pattern(CS-LBP)to realize the feature matching.Through the first three steps of SIFT: detection of local extremum,accurate positioning and distribution of orientation,the information of key points is obtained.Then it is need to establish description regions for the key points and describe them with CS-LBP to get a one-dimensional vector description.Finally,the only pair of matching points is found out by calculating the distance of the vectors.The result shows that many key points can be detected by SIFT and the time consumption is reduced because of simple calculation of CS-LBP.