Grid-based classification and spatial scale registration for efficient UAV image localization

Abstract In scenarios where Global Navigation Satellite Systems are unavailable, image-based localization methods have emerged as a promising alternative for Unmanned Aerial Vehicles (UAVs). However, current approaches often rely on homologous image matching or demand excessive computational resources, limiting their practicality. To address these limitations, we propose a Grid-based Classification Method (GCM) for UAV image localization. By gridding offline satellite images, we transform the regression problem into a classification task. Additionally, we establish a spatial image scale registration model to match heterogeneous images at different scales. To expedite geolocation, we introduce a Clustering-Quantization-based image retrieval method. Experiments conducted on an airborne computer show that our method significantly improves matching speed and accuracy, especially in large positioning areas, laying a solid foundation for efficient drone image navigation. Here, we demonstrate an average matching accuracy of 92.5% with a matching time of 53 milliseconds per frame, highlighting the practicality and performance of our approach.