Abstract:

In object detection and localization system based on local feature matching,the number of match and false match effects directly on localization accuracy. An improved SIFT matching algorithm was proposed to decrease false match and meanwhile kept the sufficient correct match number. After analyzing different match result with different match threshold in conventional SIFT feature match method, an iterative strategy for adaptive dual-threshold for image match was presented. Then a geometry constrained model based on sparse but accurate match achieved with high threshold was establish to eliminate false match in dense match set achieved with low threshold. Experimental results show that compared with other methods, the proposed method has higher match accuracy which improves the performance of object detection and localization.

Key words: feature match, scale invariant feature transform(SIFT) feature, object detection, object localization

摘要：

在基于局部特征点匹配的目标检测与定位系统中，匹配点和误匹配点的数量直接影响定位精度。为降低特征点误匹配率并保证匹配过程中有足够的匹配点数，提出了一种改进的尺度不变特征点匹配方法。分析常用特征点匹配方法中匹配结果随判断阈值变化的问题，利用循环,采用变步长的方式获取匹配图像自适应双阈值。在此基础上，利用高阈值对应的稀疏精确匹配结果建立匹配图像间的几何变换约束模型并建立约束准则，用以滤除低阈值对应的密集匹配结果中的误匹配。实验结果表明，与现有方法相比，所提方法可明显提高匹配精度，从而增强目标的检测与定位性能。

关键词: 特征匹配, SIFT特征, 目标检测, 目标定位