Accurate and Real-time Localization of Unmanned Aerial vehicles using satellite images based on Swin Transformer

Abstract

In recent years, attention has grown significantly in the unique applications of unmanned aerial vehicles (UAVs) in various fields. These applications generally require precise localization. Reliance on GPS poses issues like interference and signal loss. Recent studies suggest alternatives such as Google Earth imagery and neural networks. While promising, these methods still struggle with low accuracy and unreliable performance, partly due to limited attention to hyperparameter optimization (HPO) and feature extraction. To improve this class of methods, in this research, an accurate and fast image-based UAV localization method without using GPS is presented. The proposed method has two general stages: preflight (training) and flight (testing and localization). In the first stage, the desired flight path is taught to the drone offline. For this purpose, images of the path are sampled and then the data augmentation technique is used to augment the input data to the Swin Transformer network and increase its accuracy in the training process. In order to further improve the efficiency of the Swin Transformer network, the HPO method is used. After the internal compression by the Swin Transformer coder, an external feature extraction stage is performed based on the vector quantization method (VQM) and the Principal Component Analysis (PCA). Then the compressed vectors are stored in the drone memory. In the second stage, the drone records each live image; first it encodes it, then an external feature extraction stage is performed on the encoded vector, and localization is completed with kernel operations and comparison with memory. Simulation results show that the proposed method performs localization operations with low latency (1.65-fold reduction), reasonable memory requirements, and high accuracy (RMSE≅1.07 m).