Research On Stereo Matching Algorithm Of Binocular Vision In Complex Scene

Binocular vision as an important research direction of computer vision technology,has been widely used in aircraft navigation,ranging,virtual reality and other fields.In this thesis,we research on stereo matching algorithm,and apply it to the pedestrian detection system.On the basis of summarizing the binocular stereo matching algorithm and research status of pedestrian statistical system,this thesis describes the coordinate system of the camera imaging model,and then analyzes two kinds of binocular stereo imaging structure,and describes the principle of binocular stereo vision system calibration and correction.On the basis of the above,the constraint conditions and the similarity measure functions of binocular stereo matching and process of stereo matching are described.Finally,the advantages and disadvantages of local and global stereo matching algorithms are analyzed.Aiming at the problem of the traditional local stereo matching algorithm in dealing with the occlusion regions and near disparity discontinuity regions has high false matching rate,A stereo matching algorithm for the bilateral filtering in variable window based on mutual scale framework and refine doubtful regions iteratively is proposed.When constructing the framework of cross-scale,the Gauss down sampling and the inter-scale regularization term are introduced,and dynamically adjust the size of the cost aggregation window,so that the image information of the rough,fine scales space can better integrate.The bilateral filtering algorithm is used to calculate the cost weight in the window to improve the reliability of the matching.A doubtful regions iterative growth algorithm is proposed to fill disparity in the phase of disparity refinement,which is based on the RGB vector and gray value of the pixel points.The experimental results show that the local stereo matching algorithm proposed in this thesis can get a lower false matching rate,and further optimization can realize real-time processing.In this thesis,I introduce the idea of graph theory for the optimization of the global matching algorithm.After researching the Prim algorithm and the Kruskal algorithm,which are constructing the minimum spanning tree structure,in order to extract image features more naturally,the minimum branching tree structure based on eight adjacency is proposed based on the minimum spanning tree structure and the segmentation tree structure.It uses the gradient information of the nodes to construct a directed graph.And the segmentation of the image divides regions into low texture regions and complex texture regions.The regional aggregation criterion based on the boundary node and the root node is proposed.In the Middlebury standard test sets,the disparity map accuracy of our algorithm is better than the minimum spanning tree algorithm,and the computing speed is more prominent.On the basis of theoretical research,Pedestrian recognition and detection system based on binocular vision has been build.Extracting depth information in real-time using box filtering stereo matching algorithm based on cross-scale in the scene.Then depth stratification algorithm is used to accurately separate the depth information layer containing the character's head,which can effectively prevent the tracking chaos caused by the target adhesion.Finally,the target contours are extracted according to the morphological processing,and the object is marked in the original video stream,then the recognition and detection of the characters in the scene is finished.The experimental result shows that the target extraction algorithm based on depth stratification can accurately extract the characters in the scene,and lay a good foundation for the follow-up realization of the target tracking module and pedestrian counting module.Finally,the research work of this thesis is summarized and the further research directions are given.