Fpga-based Real-time Stereo Matching System

Unmanned vehicle, including Lunar Rover, is a kind of intelligent robot which can move and make motion decision automatically. Vision navigation is one of the most important techniques in unmanned vehicle system. The mission of vision navigation is to understand view information and recognize the target, in order to find a proper way. Stereo vision is the basic part of vision navigation. This method recovers depth of object using image matching. After that, we can do obstacle detection and 3D reconstruction. The thesis aims at study of stereo vision system for navigation and focuses on the design of real-time stereo matching.Camera model describes the projection relation between a real point and an image pixel. The target of camera calibration is to solve model parameters. A calibration method with 2D plane template is chosen according to the work environment of unmanned vehicles. Geometric rectification can be done after calibration. Rectification makes a pair of correlation points in stereopsis the same position in vertical direction. It definitely reduces the complexity of stereo matching.Among most area based matching algorithms, non-parametric method has advantage in unbalanced illumination condition. Since non-parametric stereo matching using relative intensity value instead of original one, it is more robust to illumination change. Non-parametric matching only contains simple operations such as compare exclusive-or and accumulation. For this reason, hardware logic circuit implementation is possible. The choice of matching window directly influences the match result. On the other hand, proper method of acceleration can be used in cost of large numbers of data storage.A scheme of FPGA implementation of Census stereo matching is proposed in order to meet the requirements in real-time navigation system. Design of data flow and processing modules is described in detail. Simulation is made to validate the performance of this scheme. A binocular vision sample card is introduced, which realizes the function of video sampling, signal processing and data transmission. It can be used to design and test stereo matching algorithms or other image processing methods.