Research On Binocular Stereo Vision

Mobile robot research is one of the hotspots in the robot research field. In related research of mobile robot, autonomous navigation is not only the key of realizing mobile robot to work, but also the core technology to achieve the real intelligent and autonomous move. Binocular stereo vision navigation has the advantage of wide range of detection and high navigation accuracy, and its related research subject has important theoretical significance and practical value.In this dissertation, for the purpose of navigation , beginning with the operation principle of binocular stereo vision, it is introduced how to gain the depth information from the two images which are captured from two parallel cameras. A set of low-cost and efficient hardware platform is designed according to the characteristics and requirement of the mobile robot vision navigation in order to maximize the navigation system's real-time performance and programmable flexibility. This hardware platform use high-performance digital signal processor (DSP) as its image processor, which is dedicated to video processing, and use CCD cameras to capture images. After the images are processed, the results can be displayed on a LCD to make the debug process more convenient.The camera calibration technique is introduced for the needs of generating the depth map. A linear mathematical model of camera is built and solved, the camera's internal parameters and external parameters are obtained after considering its distortion error and using non-linear optimization. Finally the spatial points are re-projected onto the image using this model and it indicates that the error is less than 0.4 pixels, which verified the accuracy of this calibration method to meet the system requirements.All the image processing algorithms are implemented on the DSP/BIOS which is an real-time operation system. After introducing the DSP/BIOS, a method of using it to obtain two images' data and maintain their synchronization is presented. Some image pre-processing and edge extraction algorithms are researched and their performance is compared. This dissertation proposes an improved algorithm of canny operator. We can gain the gradient amplitude of the sub-pixel using interpolation in neighborhood of the pixels and experiments show that the performance is optimized.This dissertation proposes an method combining the image color segmentation and edge detection algorithm for object recognition and ranging. The algorithm greatly reduces the unwanted data and improved real-time performance with ranging accuracy of 1.27%. Subsequently the algorithms of generating dense disparity map are introduced and experiments are completed to compare their stereo matching effectiveness and real-time performance, while the algorithms are transplanted on the DSP to generate dense depth map.