Research On Technology Of Binocular Stereo Vision In Obstacle Detection

Obstacle detection plays an important role in the intelligent transportation system(ITS). It is an essential prerequisite for self-driving cars. Computer vision is applied to obstacle detection, with the advantage of rich information, wide covering area and low circumstance noise. Using computer vision to detect obstacle helps to solve practical difficulties and explore fully automatic driving. Binocular stereo vision is one of the most important research aspects of computer vision. It is a way to mimic how the human visual system might deal with real scene. Therefore, with the background of obstacle detection in ITS, we have researched the key technologies of binocular stereo vision in this paper. The study mainly includes the following contents:1. Synchronization of binocular vision: Aiming at the relative motion between background and obstacles in moving intelligent vehicles, the first important step of preprocessing for binocular stereo vision is raised in this paper. From image processing angle, a synchronizing method of the two videos is proposed. Image segmentation methods are used to synchronize the two frames, which can make following stereo matching and obstacle detection more accurate.2. Camera calibration: The advantages and disadvantages of the commonly used traditional camera calibration and self-calibration methods are firstly analyzed. From both rapidity and real-time perspective, a plane calibration method of Zhang Zhengyou is used for camera calibration in this paper. Calibration experiments on collected images are carried out with MATLAB calibration tool box. And the internal and external parameters of the camera got by the experiments are verified by the projection error analysis.3. Stereo matching: After analysis and comparison, a relatively simple correspondence algorithm based on sum of absolute difference(SAD) region-matching is applied here. Moreover, in order to improve the real-time of obstacle detection in ITS, an optimized SAD algorithm is proposed from the aspect of similarity measure. The algorithm is aimed to reduce the computation in traditional SAD algorithm. Finally, with the internal and external parameters of the camera and disparity map, the depth information is extracted using triangulation. And the error analysis is conducted. The effectiveness is validated through simulation experiments on standard image pairs and real scene image pairs. And experimental results show that the optimized SAD region-matching algorithm is faster in calculation which can be better applied to actual obstacle detection.