Research And Implementation Of ARM-based Restart Random Walk Stereo Matching Algorithm

With the rapid development of computer vision in recent years, binocular stereo vision, as one of the important research direction of computer vision, its application field has been widened. Image acquisition, camera calibration, image calibration, stereo matching and 3 d reconstruction constitute the complete binocular vision system. And stereo matching, as binocular vision system's key step, the accuracy and efficiency as well as the hardware implementation of the matching algorithm directly determine the performance of the binocular vision system and its practical application. So the research of the stereo matching algorithm and its implementation is of great significance and practical value to the development of binocular vision.So far, there are many different kinds of stereo matching algorithms have been proposed, but the problems in stereo matching that weak texture, depth of discontinuity and other technical difficulties are still need to be addressed. The basic theory of the stereo matching and some good algorithms of local stereo matching algorithm is deeply studied in this paper. Based on the test and comparison of all disparity maps received from all kinds of algorithms, the random walk with restart algorithm is finally chosed to get a depth study of its aspects in need of improvement, and tried to implement on PC and ARM platform.After analysis, the random walk with restart algorithm has the problem that the edge weights is only determined by the color of the difference between pixels, and the adjacency matrix is not balanced, the two problems will increase the probability of error matching, and reduce the matching precision of the algorithm. So an improved random walk with restart algorithm is puts forward in this paper. Firstly, the edge weights is improve to be decided by the color difference and space distance between pixel collectively. Then the adjacency matrix is balanced using the double random matrix technique, to improve the matching accuracy further. Finally, the improved algorithm is tested on the test platform, and compared with the original algorithm, the results show that the matching accuracy of the improved algorithm is improved compared with the original algorithm.In this paper, the improved algorithm put forward is implemented on hardware. First, a binocular vision image acquisition platform have been set up, and the left and right camera can grabbing image of the real scenes at the same time through Matlab, and the stereo calibration is completed using the Camera Calibration Toolbox for Matlab to get the inner and outer parameters of binocular cameras, and then the real word image collected is distortion rectified by OpenC V, then using the stereo matching algorithm to get the disparity map of the real world image, the implementation of the improved algorithm on PC is completed. Then Ubuntu is built on the virtual machine, and then build the Linux+QT system on it, OpenCV is called through the QT to complete the image correction and stereo matching, then the disparity map is obtained. And then files can run on the ARM development board is generated through the cross compiler, write the Linux+QT system and working file to ARM development board, at last, the disparity map is showed on the display screen, the stereo matching algorithm is implemented on the embedded ARM so far.