Research Of Stereo Matching Technologies Based On Regional Algorithms

Computer stereo vision is the extraction of3D information from multiple digital images. Similar to the biological process stereopsis, depth information in the form of disparity of a point can be calculated by comparing its position in each image. Thus the3D information of a scene can be obtained. Therefore stereo matching is the process of searching for the image position of the same point in image pairs. This definitely is the key technology of3D reconstruction, namely, its main concern and obstacle.Stereo vision system with parallel optical axis is first studied in this paper and the shortcoming of conventional algorithms brought by fixed match window size can deteriorate the performance when processing various image features. A method with automatic match window size based on gradient is developed to solve this defect. At the same time, a parallel processing method is introduced to optimize the implementation of this algorithm that dramatically reduces redundant calculating and improves the speed and efficiency of this algorithm.For the more general case, i.e. wide baseline stereo vision systems with unparallel optical axis, a binocular vision system is chosen as the object of study. Methods of retrieving epipolar geometry in wide baseline stereo vision are studied in this paper. Since the test equipment has already utilized a more flexible and accurate approach of calibration, the rectification of stereo images can be easily processed, which simplifies the following stereo matching stage. To deal with the illumination difference and noise brought by the difference of view angles and unfavorable condition of camera, a normalized similarity measurement function is used. Like the parallel optical axis system, auto selected match window and optimized parallel processing is applied, yielding ideal test results.