A Novel Scene Matching Algorithm via Deep Learning for Vision-based UAV Absolute Localization

Almost all military systems and combat platforms rely on GNSS to achieve military missions. However, they are facing more and more intensive jamming and spoofing attacks. Due to the rapid development of computer vision technology and vision sensor, visual navigation and localization technology have gradually revealed their advantages to solve this kind of problem. In this paper, a new framework of visual localization under a GNSS-denied environment is proposed. Firstly, a novel image preprocessing method based on information entropy and granular segmentation is presented which can significantly filtrate the Google tiles in the georeferenced area to parsimony set to reduce UAV localization time. Secondly, we introduce an end-to-end image matching method based on convolutional neural network and image preprocessing to realize UAV localization which has strong robustness and localization accuracy. Finally, the proposed localization algorithm achieves experimental performance on the real images captured by DJI Phantom 3 drone. The good accuracy and fast execution time shown by experiments indicate that the proposed localization algorithm can serve as a powerful tool to keep the UAVs working continuously in the face of intensive jamming and spoofing attacks.