多视影像相对方位关系误判检测的置信传播算法  

False Detection of Relative Orientation of Multi-view Images Based on Belief Propagation

作　　者：卢俊[1] 张保明[1] 郭海涛[1] 张宏伟[1] 

机构地区：[1]信息工程大学地理空间信息学院,河南郑州450052

出　　处：《测绘学报》2015年第4期422-430,共9页Acta Geodaetica et Cartographica Sinica

基　　金：国家973计划(2012CB720000)~~

摘　　要：场景模糊或者不同场景中的重复纹理会导致影像匹配时产生大量的误匹配点,从而得到误判的相对方位关系。本文引入概率推论方法,提出了一种改进的误判相对方位关系检测算法,利用回路闭合约束构建了基于影像间相对方位关系的贝叶斯网络,推导了贝叶斯网络中的先验概率模型,并利用置信传播算法解算了贝叶斯网络中最大后验概率的求解问题。试验结果表明,利用本文提出的全局一致性约束方法可以有效检测影像间误判的相对方位关系,改善场景重建的结果,并且具有很高的计算效率。Ambiguous visual structures or repetitive textures of different scenarios would result in a large number of mismatching points in image matching stage,and false positive relative orientation is introduced which is locally consistent.In this paper,the probabilistic inference method is introduced and an improved detection algorithm of false positive relative orientations is presented.A Bayes network based on relative orientations between images is constructed using closed cycle constraints and priori probability models in Bayes network is derived and the maximum posterior probability in Bayesian network is solved using belief propagation algorithm.Experimental results show that using the proposed global consistency constraints method can effectively detect the false positive relative orientations and improve outcomes of scene reconstruction with high computational efficiency.

关 键 词：相对方位 回路约束 最大生成树 贝叶斯网络 置信传播 

分 类 号：P237[天文地球—摄影测量与遥感]