Research On Stereo Matching Algorithms

More than eighty percent of external information can be obtained by human vision.Vision is an important way for people to know something about the world. With the rapiddevelopment of technology, people try to give visual function to computers, robots, andother intelligent machines. Therefore, computer vision emerged as a new discipline.Computer vision is widely used in many fields such as drones, navigation andthree-dimensional measurement.Computer vision includes four steps: image acquisition, image calibration, stereomatching and three-dimensional reconstruction. Stereo matching aims to find thecorresponding points in two or more images taken from the same scene. The relative shiftof position of corresponding points is called disparity. With an accurate disparity map, thedepths of the points in space can be yielded via simple geometric computation, which isuseful in three-dimensional reconstruction. Stereo matching is an important problem, andthe quality of disparity map will directly affect the reconstruction results.In recent years, many researchers have been working on stereo matching. A lot ofalgorithms including feature-based algorithms and area-based algorithms are proposed.Feature-based algorithms extract local features from a stereo image pair and obtaindisparity maps by finding the corresponding features. The full range pixel correspondenceestimation is reduced to a sparse set of pixel correspondence estimation due to extractingonly local features. Therefore, the feature-based algorithms are time-saving, but usuallygenerate sparse disparity maps which are not appropriate for some applications.Area-based algorithms are classified into the local algorithms and the global algorithmsdepending on whether using global reasoning. The global algorithms have some classicalmethods, such as belief propagation and graph cut. The global algorithms are generallysummarized into a frame based on minimizing a global energy function, and the finaldisparity maps are iteratively obtained by minimizing or maximizing the energy functionusing some optimizing algorithms. The global algorithms usually can generate accurateand dense disparity maps, but are very time-consuming. The local algorithms aggregatesimilarities in a support window around each pixel. However, the process of finding theoptimal window for each pixel is a challenge. Adaptive support-weight methods proposedby Yoon in2006try to assign different support-weights to the pixels in a fixed-sizesupport window. Yoon’s algorithm outperforms other local methods and can becomparable to some global methods, which obtains much attention.Belief propagation and adaptive support-weight method are discussed in this paper.The aim is to improve some problems, such as the accuracy of depth map, the efficiencyof the algorithm, occlusions and so on. The main work and innovation are as follows:(1) Belief propagation as one of the global algorithms can obtain more accuratedisparity map than local algorithms, but is very time-consuming. To improve theperformance of the algorithm, this paper presents a stereo video matching algorithm basedon motion estimation. Firstly, the traditional BP algorithm is used to get the disparity mapof I frame. Then, for P frame, the rearranged propagating messages through referring to motion estimation information from I frame are used as the initial values for iteration ofbelief propagation algorithm. Thus this reduced the number of iterative times. Experimentresults show that the proposed stereo video matching algorithm based on motionestimation can dramatically enhance the efficiency of stereo video matching.(2) The appearances of the pixels are ambiguous because of only making use of grayinformation of the pixels in the stereo image, which causes the poor performance. Toimprove the performance, pixels are considered in the RGB vector space. Then disparitymaps can be obtained by belief propagation method in the RGB vector space.Experimental results show that the proposed method is very effective.(3) Yoon proposed the adaptive support-weight algorithm in2006. Yoon’s algorithmoutperforms other local methods and can be comparable to some global methods, whichobtains much attention in recent years. Gradient similarity is a simple, yet powerful, datadescriptor which shows robustness in stereo matching. Based on the adaptivesupport-weight approach, a matching algorithm, which uses the pixel gradient similarity,color similarity, and proximity in the RGB vector space to compute the correspondingsupport-weights and dissimilarity measurements, is proposed. The experimental results areevaluated on the Middlebury stereo benchmark, showing that our algorithm outperformsother stereo matching algorithms and the algorithm with gradient similarity can obtainbetter results in stereo matching.(4) Occlusion is one of the most challenge problems in stereo matching. In a stereoimage pair, a pixel is called occluded pixel when it only can be seen in one image. Theoccluded pixel can not obtain an accurate disparity. In order to improve the performanceof stereo matching in some regions such as occlusion region, an improved method oftrinocular stereo matching is proposed. The experimental results are evaluated on theMiddlebury stereo benchmark, showing that our algorithm outperforms binocular stereomatching and obtains an accurate disparity map.