A Study Of Binocular Vision 3D Reconstruction Technology Applied To The Intelligent Video Surveillance System

Recently the application of computer vision theories and methods to the intelligent video surveillance system has attracted an increasing attention.In this thesis,a binocular camera is used to acquire stereo image pairs,and the three-dimensional information of the target scene is restored according to a binocular camera imaging model.Difficulties in this thesis lie in how to solve the problem of light changes and noise when images are acquired by the intelligent monitoring system,and how to ensure the accuracy and real-time performance of the three-dimensional reconstruction.The main contents include the prototyping design of intelligent video surveillance system,camera calibration,stereo matching and point cloud generation.The details about this thesis are as follows:Firstly,the monocular and binocular camera models,as well as the composition and function of the intelligent video surveillance system are analyzed in detail.Then a prototype system that applies 3D reconstruction to an intelligent video surveillance system is designed.Secondly,in the phase of camera calibration,mathematical theories of camera calibration are introduced.Based on the Harris corner detection,a method of screening clustering points and non-checkerboard corners is proposed according to the unique structure of the checkerboard corner.In order to improve the position accuracy of corner detection,a sub-pixel precision calculation method is used.Results show that the proposed algorithm improves the accuracy of corner detection.Thirdly,in the phase of stereo matching,in order to meet the real-time requirements of intelligent video surveillance system and to solve the problem of mismatch caused by inconsistent exposure and noise,a new stereo matching algorithm based on AD-Census using gradient information and two-adaptive guided filtering is proposed.The algorithm uses a histogram transformation algorithm to preprocess the image,which can reduce the impact of inconsistent exposure effectively.Moreover,the AD-Census algorithm,with noise margins and gradient information,is used to calculate the cost of matching,which can reduce the effects of noise and light.Furthermore,two-level adaptive guided filtering algorithm is used to carry out cost aggregation to improve the real-time performance of the algorithm.Besides,a multi-step refinement method for disparity optimization is proposed,which can estimate the disparity value of the occlusion region and the mismatch region.The experimental results show that the proposed algorithm can get better disparity under the conditions of inconsistent exposure and noise,and it can meet the real-time requirements under GPU acceleration.Finally,in the point cloud optimization stage,the point cloud generation method and the Delaunay triangulation theory are analyzed.And a Delaunay triangulation method combining point-by-point interpolation algorithm and divide-and-conquer algorithm is proposed.Experimental results show that the point cloud obtained through this algorithm can get a visual three-dimensional model successfully after Delaunay triangulation.