基于几何约束及0-1规划的视频帧间线段特征匹配算法

针对线段因遮挡、断裂以及端点提取不准确等原因造成的线段特征匹配困难问题，特别是现有匹配算法在匹配过程中出现"多配多"时直接采取"最相似匹配"而导致丢失大量真实匹配的问题，提出了一种基于多重几何约束及0-1规划的线段特征匹配算法。首先，基于校正后视频帧间线段特征的空间相邻性计算线段匹配的初始候选集；然后，基于极线约束、单应矩阵模型约束以及点-线相邻性约束等多重几何约束，对候选集进行筛选从而剔除部分错误匹配；其次，将线段匹配问题建模为一个大规模0-1规划问题；最后，设计了一种基于分组策略的两阶段求解算法对该问题进行求解，从而实现线段特征的"一配一"精确匹配。实验结果表明，该算法与LS（Line Sigature）、LJL（Line-Junction-Line）方法相比，匹配正确率接近，但匹配线段数量分别提高了60%和11%。所提算法可以实现视频帧间的线段特征匹配，为基于线特征的视觉SLAM（Simultaneously Localization and Mapping）奠定基础。

Line segment feature matching algorithm across video frames based on geometric constraints and 0-1 programming

To deal with the problems in line segment matching due to occlusion, fragmentation and inaccurate extraction of line segment endpoints, especially when the criteria of "most similar matching" was taken in "multiple-to-multiple" matching which may lead to fateful true correspondences lost, a line segment matching algorithm based on geometric constraints and 0-1 programming was proposed. Firstly, a candidate matching set was initially computed based on the spatial adjacency after correcting the vedio frames. Secondly, the multiple geometric constraints, such as epipolar constraint, homography matrix and point-to-line adjacency, were employed to remove the false positive correspondences. Then, the matching problem was modeled into a large scale 0-1 programming. Finally, a two-stage method based on grouping strategy was designed to solve this problem, so as to realize the "one to one" exact matching of line segment feature. The experimental results show that, compared with LS (Line Sigature) and LJL (Line-Junction-Line) methods, the propsed method has a similar performance in correct matching ratio but a larger matching amount over 60% and 11%, respectively. The proposed method can fulfill the line segment matching across vedio frames, which lays the foundation for line-based visual SLAM (Simultaneously Localization and Mapping).