Some Issues Of Depth Peception And Three Dimention Reconstruction From Binocular Stereo Vision

Binocular Stereo vision is an important branch of field of Computer Vision. It simulates the visual system of the human being to perception objective world, which is widely applied into many fields, such as position detection and control of the micro-operation system, robot navigation and aerial survey, three-dimension non-contact measure,virtual reality and so on. So it has important theory value and real significant to study on some issues of depth perception and three dimension reconstruction from binocular stereo vision. Focusing on camera calibration, match strategy and matching algorithm, depth information calculation and three dimension reconstructions based on Open GL, the author had done amount of researches on these aspect. The followings are the main researching content in the paper.(1) On the basis of analysis of the traditional camera calibration methods, a semi-auto camera calibration method based on a plane plaid was proposed. The calibration method based on Tsai's two steps calibration method calculated the linear solution for a single camera. And then the solution was refined by maximum likelihood standard. Considering the radial distortion of the camera the nonlinear optimization was used to enhance the accuracy of the calibration result. The rotate matrix and translate vector of a single camera was applied into the calibration of the stereo visual system. The proposed calibration method only needed two or more images of the plane plaid from different angles. The plane plaid can move freely without any moving parameters. And the operation of the proposed method was simple but effective. Several experiment results show that the proposed calibration method is feasible and can satisfy the accuracy.(2) Aiming at matching speed and window size selecting problems of area-based matching an adaptive window size matching algorithm based on cost function of the window was proposed. In order to solve the noise and miss matching at the discontinuous region, a novel similarity measure function was designed using ρσ(n) function combined the four-direction line mask. We used error mean and variance of matching window and big window shift to judge that the disparity values of the window were same or nearly same. According with disparity distribution of the window the cost function was built and was applied to determine the window size adaptively. And integer image technique was introduced into the algorithm for enhance the matching efficient. In order to get global optimal matching for certain region block Genetic algorithm was introduced to the adaptive window matching algorithm. Several experiments results show that the proposed stereo matching algorithm could well solve the disparity discontinuity and occlusion problems. And the accuracy of matching was greatly improved by modified Genetic algorithm that is a global optimum method.(3) Generally area matching algorithm can get dense disparity. But it couldn't work well at edge. And feature matching algorithm has better robust at feature points matching except that disparity image is sparse. According to the merits and disadvantage of both area matching and feature matching a hybrid matching algorithm combined Harris corner matching and area-based matching was proposed. Considering that classic Harris corner detection algorithm was sensitive to scale changing, an improved Harris feature extraction algorithm was proposed. The improved method was on base of space and scale, which had scale invariance. The pixel at which extremum could be gotten both in the image space and scale space would be considered as a corner feature point. And the self-relative matrix M of image brightness in Harris algorithm was introduced scale adjusting adaptively for extract the corners under different scales. Considering the disparity of feature matching was sparse, we combined with area matching and constraints of disparity gradient. A dense disparity image was gained by the proposed matching scheme. Several experiments show that the proposed hybrid matching algorithm based on area and feature matching could get dense disparity image. And a good matching accuracy at disparity discontinues edge. In a word the proposed algorithm was valid and feasible.(4) Generally the depth image gotten from the form was unsmooth because of matching error and calibration bias. And some edges of the depth image were not aligned to those of the original image. Aiming at those problems a depth post-processing method was proposed, which was combined the original image and depth image to filter the depth image by three-step joint bilateral filter. The proposed method firstly used bilateral filter on the original image. And then the joint bilateral filter combined with original image was done on depth image. Finally the results of two times filters were joined together and filtered the depth image by bilateral filter again. Depth perception and measure and three dimension reconstruction are aims of stereo vision. We also have done some works on non-contact measure method of three dimensions and deduced the depth computing formula. Finally the grid building algorithm of local3D Delaunay was applied to visualize the three dimensions and depth image by OpenGL combined with MFC developing platform. Some experiments have done on both standard test data and real data. The experiments results show that the proposed post-processing method can smooth the depth image and align the image validly. And several reconstruction results also shows that the proposed calibration methods, the proposed matching method and depth post-processing method are feasible and valid.