Design And Research Of Binocular Stereo Matching Algorithms

This paper studies the binocular stereo matching in stereo vision that is for agiven scene we shot in a different perspective of two pictures. These two picturesdescribe the same scene, so the point in the scene will have a correspondencebetween different pictures. After that, we can get the spatial depth of the scene, anddraw their corresponding disparity map.Binocular stereo matching problem is thecore issue of stereo vision. This paper aims to discuss the present algorithm, andresearch the energy equation. From the perspective of the disparity space, we willgive the analysis of deficiencies of the present algorithm and design the improvedalgorithm, making eventually be able to get a border clear and high-precisiondisparity map. Finally we can provide important three-dimensional information forfuture applications such as3D reconstruction.The main works of this paper can be described as follows:1. Research on the energy equation carefully to study the stereo matchingproblem and analyze the shortage of the traditional algorithm. This paper first willdo following things: analyzing the present energy function, studying the role of eachterm, pointing out their shortcomings, and giving an improved energy function.Although there are lots of methods to solve the stereo matching, careful analysis ofthese algorithms we will find that there are several problems which are unable to bedealt with by them. Here we will disscuss each algorithm from the perspective ofdisparity space, and after describing the principle of them carefully, we will draw thediagram for each algorithm. The most important thing is that we point out theinadequacies of various algorithms, and provide a theoretical basis for the next part.2. Design the shortest path SP algorithm to solve the stereo matching problem,so that it can get a high-precision disparity map with a less complexity. Studing fromthe disparity space, we find the inadequate for each algorithm, such as the localoptimization led to the matching accuracy is low, or when considering the globaloptimum the complexity is too high, or too many iterations. Here we make someimprovements on the present stereo matching algorithm and also design a newsolution model. Now the problem is mapped to solving the shortest path. Through the conversion, we make the solution of the energy function into the di stancebetween two points, and here the shortest distance is the optimal solution of theenergy equation, so form the solution we can get the disparity for each pixels.3. Design the RBFSM based on radial basis function neural network algorithmto focus on non-standard case of the stereo matching problem. The accuracy ofstereo matching algorithms have been greatly improved now, however, carefulanalysis found that they just considered standard images which are corrected before.When pictures are not in a standard condition, the effect of each algorithm will besignificantly reduced, and that means the robustness of these algorithms are poor.Here we design a new stereo matching algorithm based on neural networks to studythe stereo matching when pictures in the case of non-standard, such as pictures havea quit different brightness or one of pictures has a slightly rotation, in order toextend the field of application. Experiments show that the new algorithm has astrong robustness, and also have a good effect on the actual scene, so we can also geta satisfied disparity map and accuracy in the end.