Research On Perception Technology For 3D Target Based On Binocular Vision

Binocular vision system aims at obtaining 3D structure information and position information of target.It is reliable in processing scenery by simulating Human Vision System,and is widely applied in many fields,including intelligent driving,virtual reality,aerospace and 3D non-contact measurement.A complete binocular vision system includes image acquisition,camera calibration,image correction,stereo matching and 3D reconstruction.Among them,stereo matching is the focus and difficulty of vision research because the matching results are influenced by many factors,such as illumination,smoothness of object surface and occlusion between objects.Although many stereo matching algorithms have been proposed,there are still some deficiencies in accuracy and efficiency.In this thesis,a complete binocular vision system for perceiving 3D objects is studied and implemented.Aiming at some problems existing in current stereo matching algorithms,corresponding improved algorithms are proposed.The main research work of this paper is as follows:1.The imaging model of the camera and the four coordinate systems involved are introduced.Then method for camera calibration is studied,and the Zhang Zhengyou method is applied to accomplish the calibration experiment.The camera's internal and external parameters and distortion parameters are obtained,and then the image is corrected according to these parameters to eliminate the distortion in the image and make the polar line aligned.2.The stereo matching algorithm based on bilateral filtering,which has excellent performance in local algorithms,is deeply developed.It considers the Euclidean distance and gray-scale distance between the pixels in the support window and the central pixels,but the gray-scale difference is only a prior knowledge of the depth difference.Thus mismatching is easy to occur when dealing with areas with large gray changes and gentle depth changes.To solve this problem,a bilateral filtering cost aggregation method based on depth information is proposed,which can be used for parallax refinement.The matching accuracy is improved by 7.4%,but the parallax refinement is time-consuming.Therefore,the idea of multi-scale is introduced to construct an image pyramid to iteratively obtain the disparity map.The experimental results show that the accuracy of the algorithm is improved by 8.7% when the running time is only increased by 1.66%.3.An improved semi-global stereo matching algorithm based on segment-tree is proposed.In semi-global matching algorithm,there is a problem of incomplete path in the smoothness constraint which is along straight line.Therefore it is easy to produce mismatching problem in weak texture region.For this problem,an improved algorithm is proposed.A segment-tree is constructed over graph,and a rough disparity map is obtained by aggregating cost on the segment-tree which is as the dynamic programming path.Then the segment-tree is reconstructed with the rough disparity map,and the final disparity map is obtained by aggregating cost on the new segment-tree.In the process of cost aggregation,it just needs to traverse the segment-tree twice to complete all calculations of aggregated cost.Numerical experiment results show that our method solves the problem of incomplete path well.The average error rate of proposed algorithm is 45.36% lower than that of the semi-global matching algorithm in non-occluded area,and the running time is reduced by 50%.4.A dense 3D reconstruction method based on disparity map is investigated.According to the principle of triangulation,the 3D point cloud data of target is obtained.The characteristics of surface reconstruction a lgorithm are analyzed,and the surface reconstruction method based on Delaunay triangulation is applied to mesh the 3D point cloud data.Then the texture mapping of the mesh model is carried out to obtain the 3D model with texture information.