Location Detection of Unmanned Aerial Vehicles with Terrain-Based Navigation Using Autoencoder Deep Neural Network

Global Navigation Satellite Systems (GNSS), primarily GPS, provide critical positioning information for unmanned aerial vehicle (UAV) navigation. However, GNSS signals can be obstructed, weakened, or intentionally jammed in certain environments, such as canyons and dense urban areas. This adversely affects the safe and autonomous flight capabilities of UAVs. This study aims to investigate the effectiveness of using autoencoder neural networks for UAV navigation in complex environmental conditions, without relying on GNSS signals. Autoencoder neural networks are highly successful in learning complex representations from visual data, and thus can be utilized for UAV environment perception and position estimation. The study created a dataset consisting of Google Maps data and nadir images captured by the UAV. The autoencoder neural network was trained using the Google Maps dataset. The trained neural network-based algorithm processes the nadir image captured by the UAV and can locate it on a pre-processed map with high accuracy using template matching algorithms. The findings of this study offer a novel approach for GNSS-independent UAV navigation. Autoencoder neural network-based navigation algorithms can significantly enhance the potential of this technology by serving as a reliable and effective backup navigation system against the possibility of GNSS signal degradation in more complex missions. Future research can focus on further improving the performance of autoencoder neural network-based UAV navigation. In this context, the use of high-resolution digital elevation model maps, training with more comprehensive and accurately labeled datasets, and better learning of visual features can facilitate the development of a more robust navigation system. Additionally, integrating methods such as the Kalman filter can reduce the impact of sensor errors and environmental uncertainties, thereby enhancing the accuracy and reliability of position estimates. Pyramid data structures containing maps at different spatial resolutions can also enable autoencoders to process visual information at different scales, potentially improving positioning performance. Researching and developing these elements has the potential to significantly enhance the success of autoencoders in UAV navigation. Furthermore, testing these additional factors in real-world conditions and integrating them into practical applications will contribute to the maturation and increased reliability of the technology. Keywords: Autoencoder Neural Networks, Visual Navigation, Computer Vision, Image Processing, Template Matching Algorithms