Research On Spatial Depth Information Recoery Method Based On Binocular Stereo Vision

Real perception of the objective world is an eternal topic that humans are constantly pursuing.Binocular stereo vision plays an important role in human life,because of its similar structure with human visual perception.Therefore,it is very important to restore this scene's 3D coordinates in objective world from binocular images which are collected by binocular vision system.This thesis discusses the issues of camera calibration experiments,feature description extraction and matching in binocular images,the generation of parallax,and the solution to third-dimensional coordinates.The main work of the thesis consists of the following aspects:First,the camera imaging model is classified and the coordinate transformation relationship is deduced.Direct linear transformation method,Tsai two-step method and Zhang Zhengyou camera calibration method are introduced.Then Zhang's calibration method is used to calculate the internal and external parameters of binocular camera according to the conversion relationship between world coordinates and pixel coordinates.The error is analyzed by re-projecting the feature points,and results shows that the pixel error is less than 0.1 pixels,verifying the reliability of Zhang's calibration method.Second,the binocular image is filtered and denoised.The contrast results of mean filtering,Gauss filtering,bilateral filtering and median filtering show that median filtering not only can reduce step noise but also preserve edges and details of the image,which is the most suitable filtering method in this thesis.The stereo rectification method of Bouguet is introduced,and used to correct the denoised image to parallel.The result shows that the binocular images have a striking correction effect which have the same line number.Then,a method is proposed which is to match the Harris corner points by using the statistical autocorrelation coefficient,and it is compared with the SIFT and SURF algorithms in terms of the number of feature points.At the same time,the number of matching points and running time is recorded.The results show that the method proposed in this paper has relatively high accuracy,but the SURF algorithm is the best choice considering the integration time and the number of matching points.Finally,a parallax generation method is introduced which combines the SURF and GC method.This method is validated by comparing the local parallax generation method,the SGM method and the proposed method.The scene's 3D coordinates in this image is calculated by using calibrated camera parameters and generated disparity map.Then the 3D point cloud can be reconstructed.The final result verifies the effectiveness of the camera calibration,image matching algorithm and parallax generation method.