Research On Binocular Matching Algorithm Using Fisheye Lens

With the rapid development of computer vision technology, more and more attention ispaid to stereoscopic vision and panoramic vision. Using two or more cameras shootingsimultaneously, stereo vision could obtain three-dimensional information of the scenes.Comparing with the planar image, one-dimensional depth information is plused.Thereforestereo vision reflects the orientation information more realistically, in order to have a widerange of applications in many fields.While panoramic vision which could capture a360-degree image by multiple cameras in onetime, is widely used in monitoring and robotvision, etc. Thus a new solution is provided for obtaining the real-time panoramic image ofthe scenes. Therefore, using fisheye lens with large field angle to achieve panoramic imagingand using binocular lenses to achieve stereo vision, in order to build a binocular stereo visionsystem. And binocular matching algorithm of fisheye images is defined as the research topicin this paper. It is aimed to provide a basis for three-dimensional reconstruction based onfisheye images and fisheye images mosaic technology. Two solutions are proposed againstthe research topic, that fisheye image matching algorithm based on corrected images andfisheye image matching algorithm based on uncorrected images. Discussion is commencedfrom the following aspects.Firstly, the background and significance of the research, as well as the domestic andforeign development status and problems are presented. And a clear solution is determined inthis paper. Secondly, the fisheye lens imaging principle, stereo vision theory, polar correctiontheory, the general procedure of planar binocular matching algorithm and existing fisheyeimage matching method based on uncorrected images are introduced. Then, on the basis ofthe knowledge of the stereo matching technology and basic theory of the fisheye lens, the twosolutions are analyzed theoretically and verified the feasibility by experiments.For the fisheye image matching algorithm based on corrected images, fisheye lensdistortion coefficients are obtained using parameters separation method and effective area isextracted firstly. Then a fast matching algorithm based on segmentation is raised. For thisalgorithm, image segmentation is conducted firstly using MeanShift algorithms. Then theinitial disparity map is calculated by rapid cost superposition strategy. Next, disparity plane isfitted and parallax region is merged. And energy function is defined by collaborativeoptimization algorithm to optimize the parameters of planar. Finally calculate the finaldisparity map. Experiments results show that the final disparity map can reflect the positionalrelationship of the scenes well and the algorithm is of good Real-time performance. However,the error introduced by correction needs to be further reduced.For the fisheye image matching algorithm based on uncorrected images, the initialmatching pair and its affine matrix are obtained by SIFT algorithm firstly. And sorting by the NCC scores of the matching pairs and their affine matrixes, in order to determine the seedmatching pair and standardize corresponding neighborhood. Then make diffusion on thenormalized neighborhood to get a new matching pair and its affine matrix throughneighborhood transformation. Repeat the above steps until no unmatching points. Finally,eliminate the abnormal points and obtain the final quasi dense disparity map. Theexperimental results show that this algorithm can get more accurate quasi dense disparity mapwith relatively long time consumption.