基于视差点的大遮挡检测和立体匹配方法

Stereo Matching and Large Occlusion Detection Based on Disparity Points

提出一种基于视差点的方法来解决在高质量立体视差图生成过程中所出现的遮挡问题.首先证明同名核线对应的视差函数曲线可近似为一条分段直线,然后在此基础上引出视差点的概念.在视差点的结构中利用两个参数分别描述左右遮挡量,使得所提出的方法能够很好地解决遮挡问题.通过分析视差点及其邻域的灰度特性,提出一种分层假设证实和Marquardt-Levenberg(M-L)算法相结合的方法从同名核线图像中提取出候选视差点,然后采用不定期的动态规划(dynamic programming,简称DP)算法获得核线最优的视差函数.利用国际标准数据对提出的方法进行了测试,并与其他方法作比较,实验结果表明,它的匹配效果是目前核线最优方法中最好的,仅差于几种优秀的全局最优方法,但其计算复杂度要远低于全局的方法.

An algorithm based on disparity points to solve the occlusion problem in the process of building high-quality stereo disparity map is presented in this paper. It is firstly proved that the disparity curve corresponding to a pair of epipolar-line images may be approximated by a group of piece-wise straight lines, and then the definition of disparity point is introduced. In the parameterization of a disparity point, two parameters are used to describe left and right occlusions so that the occlusion problem can be successfully solved in the approach. By analyzing intensity property of a disparity point and its neighbor points, an approach which combines stepwise hypothesis-verification strategy and Marquardt-Levenberg (M-L) algorithm is devised to extract the candidate disparity points from the epipolar images, and then aperiodic dynamic programming is employed to search the epipolar-optimal disparity function. The proposed method is tested by using the international standard image data and compared with other methods, and the experimental results show that its performance is the best among epipolar-optimal methods and worse than some excellent global-optimal approaches, but its complexity is much lower than the global-optimal approaches.