Development of deep learning-based autonomous landing system for unmanned aerial vehicles

The rapid development of artificial intelligence applications has brought significant innovations, particularly in the field of UAVs and autonomous flight algorithms. However, companies working on UAV technology face challenges in using UAVs for delivery due to collision issues during approach and landing. These problems stem from the difficulty of landing in crowded or complex environments. This thesis presents a new approach that uses computer vision and deep learning techniques to accurately and in real-time identify landing sites. For this purpose, an environment with numerous obstacles and objects was designed in the WEBOTS robot simulation environment, and a Mavic 2 Pro UAV was used. The YOLOv8 algorithm was trained to recognize images of the landing pad taken from real urban environments, the internet, and the proposed simulation environment. The trained model was integrated into the UAV's control code. This enabled the UAV to autonomously recognize the landing pad in a complex environment and land accurately based on the algorithm and computer vision techniques. The study achieved successful results in determining the landing area with more than 95% accuracy in most experiments conducted in the simulated environment. This thesis offers an approach that can contribute to solving many obstacles and challenges faced in autonomous flight and is presented as a suitable solution, particularly for real-world applications.