一种多视角自适应的模板匹配目标检测方法

A Multi-view Adaptive Template Matching Method for Object Detection

基于灰度的模板匹配方法难以解决航空图像多视角变换问题,而基于特征点的模板匹配方法难以解决低分辨率小目标模板图像匹配不稳定的问题。为此,提出了一种基于变换矩阵空间优化搜索的模板匹配方法。首先将多视角下的投影变换空间进行离散化建模,利用归一化灰度的模板与实时图像,以投影后模板图像与实时图像之间的绝对误差和(SAD)建立优化模型;然后通过优化搜索算法寻找到模板图像与实时图像之间的最优变换矩阵,检测出实时图像中的包含的模板目标;最后针对搜索的时间复杂度较高问题设计了基于分支界限法的加速算法。利用公开数据集和实际图像进行仿真实验,结果表明所提的模板匹配方法相比传统特征匹配方法对于高斯噪声、高斯模糊和图像有损压缩等图像退化具有更好的适应性,在大视角差异和低分辨率条件下具有更低的投影误差和更高的稳定性,并解决了单模板多目标的匹配检测问题。

Gray-level-based template matching method is difficult to solve the problem of projective variations due to multiple viewpoints of images,and feature-based matching method can＇t yet handle the template image of small target with low resolution. Aiming at above problems,a template matching method based on optimization searching in projective transform matrix space is proposed. Firstly,the optimization search space is established by discretizing the real projective space of multiple views.Secondly,an optimization model is created with the sum of absolute difference（SAD） of normalized template and real-time images.Thirdly,an optimization searching algorithm is designed for finding the optimal projective matrix between template and real-time images,which means that the targets in the real-time images corresponding to the template have been detected.Finally,for the high time complexity of the searching algorithm,a new algorithm based on branch and boundary is designed to accelerate the optimization searching. Experiments on VOC 2010 data-set and captured images of actual complex scene demonstrate that the proposed method performs more adaptively than SIFT-based matching method on degradations of images,such as Gaussian blurring,Gaussian noise and JPEG compression,and also more robust with lower projective error for images matching with low resolution and great difference of multiple viewpoints. Particularly,the proposed method perfectly solves matching of single template with images of multi-target.