针对弹体低空飞越沙漠、山地等区域时,实时图与基准图之间尺度差异大,匹配误差大的问题,提出一种用改进的形状上下文结合蚁群算法对边缘特征进行部分形状匹配的方法。为克服异源图像边缘特征的形状差异,在形状上下文算法的基础上,提取边缘邻域特征,根据边缘邻域特征相似性进行粗匹配选出候选边缘,再利用候选边缘和待配准边缘点的位置关系和形状相似度使初始化后的蚂蚁遍历所有边缘点,然后利用转换后匹配点对的距离加权和计算信息素增量,由相邻匹配点对的方向约束关系计算其权重,并根据相邻匹配对的距离关系辨别出没有正确匹配关系的边缘片段,则剔除片段上的匹配点对,对应匹配对的权重为零,最后多只蚂蚁共同多次迭代得到最优匹配结果。在实测红外可见光数据集上进行景象匹配实验,结果表明,提出的算法显著优于SIFT算法,能够克服异源图像上边缘特征的形状差异,实现两个边缘特征的精确匹配,定位精度比传统的边缘特征匹配方法提高了43.71%。

A method is proposed to partially match edge features using an improved shape context combined with ant colony algorithm to address the problem of larger scale differences and matching errors between real-time and baseline images when projectiles fly over deserts, mountains, and other areas at low altitude. To overcome the shape differences of edge features in heterogeneous images, edge neighborhood features are extracted based on the shape context algorithm. Rough matching is performed based on similarity of the edge neighborhood features to select candidate edges. Then, the position relationship and shape similarity between the candidate edges and the quasi-edge points to be matched are used to make the ants traverse all the edge points after initialization. Then, the transformed matching point pairs' distance weighted sum is used to calculate the pheromone increment, its weight is calculated from the direction constraint relationship of adjacent matching point pairs. And based on the distance relationship between adjacent matching pairs, the edge segments that do not have the correct matching relationship can be identified, when removing the matching point pairs on this segments, the weight of the corresponding matching pairs is zero. Finally, multiple iteration by multiple ants can obtain optimal matching result. Scene matching experiments are conducted on a measured infrared visible light dataset, and the results showed that the proposed algorithm is significantly better than the SIFT algorithm and can overcome the shape differences of edge features on heterogeneous images, achieve accurate matching of two edge features, and improve localization accuracy by 43.71% compared to traditional edge feature matching methods.