Extraction Of Scene Depth Information Basing On Double Horizontally-paralleled Viewpoints

Recently,the computer vision field has showed increasing interests in the extraction methods of scene depth information based on horizontally-paralleled double viewpoints.The method mainly works on how to create the 2.5D geometric information of certain scene through multiple images with the same size under the same scene.It's the most fundamental and key problem to computer vision field and non-touched measurement research.And it is based on the imaging principle of human eyes.The principle of this method is that the depth information is obtained from the horizontal disparity of correspondent pixels in the image pair.It supplies the key information to robot navigation, image-based rendering,object recognition,image segmentation and video processing etc.For the practical applications,stereo matching is not recommended from the point of mathematical view considering the common existence of occlusion and noise.Researchers have introduced various algorithms to find out a solution and already received a satis fying effect.This paper proposes a fast stereo matching algorithm for obtaining the depth map effe ctiv ely.A 3D model is used to store and compute the depth map.And the optim ized windows area-based stereo algorithm can fastly generate the initial matching value.Uniqueness and contin uity,presented by Marr and Poggio,are two assumptio ns when increasing the matching correspondence.It means for each pixel in the image pair there exists only one corresponding depth value in the depth information map.And in most area,the depth value of neighbouring pixels is continous.The selection of depth value is based on the restriction method proposed in this paper.We restrict the depth value using inhibition area,and then keep the details information with the intial matching result as the coeffic ient.Experimental results show that the fast algorithm in this paper has advantages on the accuracy of depth value,comparing to the famous fast SMP algorithm.Also,horizontal and vertical slant is fully discussed and analyzed is this paper.First,it is demonstrated that correspondence and segmentation are chicken-and-egg problems,which can be only solved together.We then introduce horizontal and vertical slant to make a first order approximation to piecewise contin uity.Since the horizontal slant surface is horizontally stretched,N pixels on the scanline in one image may correspond to M pixels on the same scanline in the other image while N is not equal to M.This requires three important modifications on the existing stereo algorithm:(a) modify the existing intensity matching method due to unequal sampling,(b) unequal number of pixels in the image pair must correspond to each other,(c) the traditional uniqueness constraint for detecting occlusions need to be modified to the new interval uniqueness constraint.At the end of this paper,the experimental results prove the correctness of this method.