A new geographic positioning method based on horizon image retrieval

In the wild, the positioning method based on Global Navigation Satellite System (GNSS) can easily become invalid in some cases. We propose a geo-location method that can be used without manual feedback even in the absence of GNSS signals. This method belongs to a vision-based method, which is realized through horizon image retrieval. Horizon image retrieval is a task with a huge database in which each image has a unique label, and different images cannot be divided into a single category. To solve this problem, we develop a new training method called “a few-shot image classification training method for serving image retrieval problems” (FSCSR). This method involves training on multiple few-shot classification tasks and updating the parameters by testing on image retrieval tasks, thereby obtaining a feature extraction model that meets the retrieval requirements. A new neural network, named HorizonSegNet, specifically designed for horizon images is also proposed. HorizonSegNet, trained with FSCSR, demonstrated its effectiveness in the experiments. Besides, a search strategy called “area hierarchy search” is proposed to increase the accuracy and speed of retrieval as well. In the experiments that conducted on 182.72 km² of land, our positioning method achieved a 95.775% success rate with an evaluation error of 40.23 m. The results verified the conclusion that our positioning accuracy is generally higher than that of other positioning methods. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.