Research Of Stereo Matching Algorithm In Binocular Vision

Binocular vision in machine vision has received more and more attention from people,since it can simulate the process of the human eye's perception of the world and perform a three-dimensional reconstruction of the surroundings,it has a wide range of applications in areas such as driverless cars,robot navigation and augmented reality.Stereo matching,the most important and challenging part of the field of binocular vision system,has a great influence on the three-dimensional reconstruction.In order to improve the matching accuracy and time efficiency of stereo matching algorithm,the thesis does research on the stereo matching algorithm based on the difficulties and essential problems of stereo matching.There are a large number of disparity discontinuities and low-texture regions in the real scene,which makes stereo matching difficult,and many stereo matching algorithms do not work well in these regions.So this thesis puts forward a matching algorithm called Local Stereo Matching Algorithm Based on Improved Census Transform and Multi-Scale Spatial,aiming to improve the matching results for those regions.Firstly,by analyzing the limitations of the traditional Census transform in the cost calculation,an improved Census transform method and self-adaptation of weights based on pixel's mean information of window is proposed.In cost aggregation step,guided map filtering algorithm with excellent gradient preserving property is introduced into gaussian pyramid structure and regularization is added to strengthen cost volume consistency.In the disparity selecting step,a series of optimization methods,such as the outlier detection,the region voting and the sub-pixel enhancement,are used to improve the correct rate of the disparity map.The algorithm is tested on the basis of Middlebury benchmarks,and the results demonstrate that the algorithm exhibits good performance and is efficient for disparity discontinuities and low-texture regions.The essence of the stereo matching problem is to find the matching points corresponding to the left and right images,and it can be regarded as the process of image feature analysis and key point extraction.While the convolutional neural network has a strong capability of feature extraction,which can gain more complex non-linear relationship.Thus,this thesis proposes a stereo matching algorithm based on convolutional neural network to solve matching problems.Two convolutional neural sub-networks are used to extract the features of matched image pairs,and the output cascade is input to the full convolution layer to complete the matching cost calculation,which contributes to better matching cost results compared with traditional stereo matching methods.Furthermore,the employment of the method of semi-global matching and the construction of global energy equation are involved to optimize the matching cost.By demonstrating the performance of the proposed algorithm on KITTI benchmarks,the results show that our algorithm outperforms many state-of-the-art methods and confirms the effectiveness of approach.In addition,the application of stereo matching algorithm in 3D reconstruction is studied.To put it explicit,this thesis analyzes the relationship between disparity map and depth information,employs Delaunay triangulation method to make the world's three-dimensional space point set coordinates grid topological structure,and aims to achieve the visualization of three-dimensional scene on the basis of the OpenGL platform.This thesis does research on the stereo matching algorithm in binocular vision,two stereo matching algorithms with higher matching accuracy are proposed and the 3D reconstruction based on the disparity map is also studied,which illustrate the importance of stereo matching algorithm and the practicality of the content of the thesis.