面向无人机绝对定位的遥感影像快速检索方法

Fast retrieval method of remote sensing image for UAV absolute location

针对在复杂环境下无人机景象匹配导航中的视觉绝对定位问题,提出了一种聚合深度学习特征的实时影像快速检索方法。首先,引入可训练软分配深度学习框架—NetVLAD,结合VGG16网络提取并聚合生成影像稳定的全局特征表达向量;其次,在初始检索阶段,使用KD树结构对影像全局特征向量构建检索索引,在不损失检索精度的前提下提高检索速度;最后,使用皮尔逊积矩相关系数对初始检索结果进行快速预判断,自动过滤初始检索结果,对于需要重排序的影像则采用特征学习匹配算法——图神经网络SuperGlue进行匹配重排序。所提方法在公开的夏季和冬季遥感影像数据集分组进行实验,实验结果表明:未重排序条件下,初始检索结果第一张影像平均准确率达到了58.27%,部分特征较好地区准确率达到了85%,对不同时相遥感影像也有很好的适应性,平均检索一张影像耗时3.7 s,可为无人机景象匹配导航的初始定位提供参考。

Focusing on the absolute positioning problem of UAV scene matching visual navigation in complex environment,a fast real-time image retrieval method based on aggregation of deep learning features is proposed.Firstly,NetVLAD,a trainable soft assignment deep learning framework,is introduced to extract and aggregate the image stable global feature representation vector with VGG16 network.Secondly,in the initial retrieval stage,KD tree structure is utilized to construct the retrieval index of image global feature vector,which can improve the retrieval speed without losing the retrieval accuracy.Finally,the initial retrieval results are judged quickly by using the Pearson product-moment correlation coefficient that can automatically filter the initial retrieval results.Graph neural network SuperGlue,a feature learning and matching algorithm is utilized to match and reorder the images that need to be reordered.The proposed method is tested by grouping open summer and winter remote sensing image datasets.The experimental results show that under the condition of no reordering,the average accuracy of the first image of the initial retrieval results reaches 58.27%,and the accuracy of some areas with better features reaches 85%.It also has good adaptability to remote sensing images of different time phases and takes 3.7 s on average to retrieve an image,which can provide reference for UAV scene matching navigation initial positioning.