大视角下非平面场景的图像特征匹配  

Image matching for non-planar scenes under wide baseline views

作　　者：刘坤[1] 罗予频[1] 杨士元[1] 

机构地区：[1]清华大学自动化系,北京100084

出　　处：《清华大学学报（自然科学版）》2010年第4期499-502,共4页Journal of Tsinghua University(Science and Technology)

摘　　要：为了改进了大视角下非平面场景的图像匹配质量,提出了一种特征匹配传播算法。传统算法的隐含假设是图像中以任意一点为中心的区域均满足局部共面性条件,该算法突破了这一假设条件的限制。通过对图像局部区域共面性条件的判断,在给定匹配点周围利用局部仿射变换和全局极线几何约束对匹配点进行扩充,在两幅图像的共同可见区域内产生准稠密的匹配点对。匹配可以自动传播到局部变换模型与初始变换不同的区域。通过在真实图像上的验证表明,该算法有效提高了在大视角间隔和透视情况下的图像匹配质量,增加了匹配数目。A match propagation algorithm was developed to improve the image match quality for non-planar scenes under wide baseline views. The traditional algorithm is based on the implicit assumption that the region centered on any pixel in an image satisfies the local coplanar situation. The algorithm breaks through the traditional assumption. After determining the co-planar conditions of local regions around the initial matches,the local affine transformation and constraints of global epipolar geometry are used to expand new matches to obtain the quasi-dense matches. The matches can be propagated into the regions whose local transformation is different from the initial one. Experiments show that the algorithm can improve the matching quality and increase the matching number under wide baseline and non-planar scenes.

关 键 词：特征匹配 大视角 非平面场景 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]