Research On A Depth Information Estimation Algorithm That Feature-matching Combined With Region-matching

With the development of computer technology, stereo vision has receivedwidespread attention and application, and become a hot topic. The most importanttechnology of stereo vision is to reconstruct a three-dimensional scene by using thepoint cloud which is acquired from the real scene images. As a hot issue of stereovision, obtaining depth information is widely applied in the fields of military, medical,film and television, for instance,in deep-sea exploration, automatic barrier UAVs orsatellites identification.This subject can solve a lot of substantive issues.To acquire the dense three-dimensional point cloud which is for reconstruction,after investigated SIFT (Scale-invariant feature transform) and region matchingalgorithm, this paper puts forward a method combined SIFT with region matchingalgorithm and optimized.First of all, using SIFT feature matching to obtain thedisparity of images, and mark the feature points.Then, using the region matching todeal with the area of non-feature points which under the constraint of featurematching results. If match template exists the feature points, the matching point’srelative position can acquire by the distance of the feature points and results ofmatching point information,which could reduce the calculation of searching.Otherwise, by using the value of the disparity gradient and adjacent point of parallaxestimates search scope.Finally, in order to shorten the operation time, put forward thecoordinated mechanism of CPU and GPU, using OpenCL API to make the searchrange estimate calculation into executable program objects which running by GPU, inorder to save time of scanning match template. The experiments show that the CPUand GPU collaborative mechanism save the operation time.The images of experiment are from stereo image database which is published bythe American Middlebury College. Theresults show that compared with SIFT andregion matching algorithm, the method of this paper reduces computing time andimproves the matching accuracy, and the disparity map is dense.