A Non-GPS Return to Home Algorithm for Drones using Convolutional Neural Network

Abstract

The increasing vulnerability of Unmanned Aerial Vehicles (UAVs) in both military and civilian applications to Global Positioning System (GPS) spoofing attacks poses significant threats to security and safety like hijacking, collision, and potentially human casualties. Despite extensive research on countermeasures, existing solutions remain ineffective, as they rely on GPS data that is often the target of the spoofing to Return to Home (RTH) or the availability of ground sensors. This article proposes a drone’s RTH mechanism based on non-GPS data utilizing Aerial Images and Convolutional Neural Network (CNN). The drone, as it flies, collects frames and the moving directions (degrees) to use it later for training the CNN model that will enable the drone to autonomously navigates back to its homebase (RTH). Several experiments have been conducted using the proposed method and it demonstrates promising results. The average distance of RTH distance to home base is 20 to 40 meters using 50 epochs only. The Mean Absolute Error (MAE) on the converted degrees (Cosine and Sin) reached below 0.02 during training. The findings not only offer a viable solution to the GPS spoofing problem but also significantly enhance the drone’s RTH reliability and improve the robustness of the drone’s ability to RTH.