Cross-view UAV image matching and localization using deep convolution features

Abstract

Matching the UAV image with the airborne reference image containing geographic information can achieve the precise positioning of the UAV image target. However, there are significant differences in the imaging mechanism, image perspective and scale between UAV images and satellite maps. In view of the above reasons that lead to low image matching accuracy and large positioning error, this paper realizes an image matching algorithm based on deep convolution feature. Specifically, the multi-scale feature descriptor is constructed by using the feature maps output by different layers of convolution network, and then the feature point matching is realized based on a dynamic interior point selection method. The five shooting perspectives of the same target image of UAV are divided into vertical reference image and 45° inclination angle for matching experiments, and the performance differences between the proposed algorithm and the traditional method are compared. Experiments show that it has good positioning accuracy and better anti-angle change ability than traditional methods.