Abstract: According to image feature space of time-varying parameters，the stereo matching problem is introduced.Stereo vision matching algorithm is usually made for time-invariant.Therefore，for the time-varying probability distribution of the image feature space，these methods can not effectively realize the matching of dynamic three-dimensional vision.To solve this problem，this paper builds a Dynamic Bayesian Network（DBN） for the edge of Graph Cuts（GC）.The unsupervised Bayesian learner，classifier and DBN are syncretized.It will acquire the rules that the probability or stochastic process model evolves with time so as to the image feature space of time-varying parameters can be estimated dynamically and accurately.Model results are used as the corresponding inter-element.In the edge region，for Belief Propagation（BP），it realizes the matching of dynamic three-dimensional visual images.

Key words: belief propagation, dynamic Bayesian network, EM algorithm, Bayesian learning, Schwarz information criterion

摘要： 提出图像特征空间概率分布参数时变的立体视觉匹配问题。立体视觉匹配算法通常是针对时不变提出的。因此，对于概率分布时变的图像特征空间而言，这些方法均不能有效地实现动态立体视觉匹配。针对这一问题，提出的方法是对图像分割边缘构建动态贝叶斯网，结合贝叶斯学习并充分利用其学习所获得的图像空间概率模型变化演进的规律，得到较准确、平滑地图分割的动态结果，以此作为基元间的对应性，在边缘区域进行置信传播，实现图像的动态立体视觉匹配。

关键词: 置信传递立体匹配, 动态Bayes网, EM算法, Bayes学习, Schwarz信息准则