Embedded System Of Binocular Vision And3d Reconstruction Technique Reserch

The binocular stereo vision technique is becoming a hot research topic and main content in the computer vision domain, using human visual system simulating and relevant technology of computer vision. This method is based on shotting the scene from different angles using binocular cameras and reconstructs a3-D model, which is widely used in scene reconstruction, robot navigation, object recognition, precision industrial measurement, virtual reality and other fields.The key technology of binocular stereo vision including images acquisition, camera calibration, image rectification, stereo matching and3D reconstruction etc.. Nowadays, most of vision platform is only responsible for image capturing, which need computer software to accomplish stereo matching algorithm. In order to solve this problem, this research based on Census matching algorithm and using FPGA chip to capture image, design an embedded system platform. Moreover this paper proposes fusion of local adaptive support-weight approach and belief propagation for stereo matching. The feasibility and high accuracy of this method has been verified by simulation on computer system, and realizing the scene3D point cloud reconstruction according to the match results. The work of this paper as follow:(1) Based on FPAG designed binocular vision acquisition and display system platform. The system use the Cyclone II series FPGA as the main processing chip, extended with binocular video input daughter card and the USB daughter card. First capturing video and images from the two CMOS cameras, frames storing and switch controlling. Then alternate shows on the monitor through the VGA interface. The system realization of time sharing and stereo display technology, then transfer the image data to the computer software through USB at the same time.(2) In the respect of camera calibration, using matlab calibration toolbox to calibrate two cameras respectively, getting internal and external parameters. Then make stereo matching algorithm searches from two dimensional to one dimensional by the Rectification of geometry relationship.(3) An algorithm fusion of local adaptive support-weight approach and belief propagation is proposed to resolve the inaccuracy among the depth discontinues area. Firstly, using improved local adaptive support-weight approach to get initial disparity, and using cross-check technology to detect incredible pixels. Secondly, transfer the segmented image adopting new strategy of unsymmetrical message propagating technology.Finally, correcting incredible pixels. Experiment results indicate that the algorithm has a better performance in low texture and depth discontinues region. In order to get the real time depth image, based on the FPGA experimental system realizing Census stereo match algorithm, compared to the similar method, the resource used and the speed have greatly improved.(4) According to principles of stereo vision computing3D coordinates, using OpenGL realizing the scene3D point cloud reconstruction.