Development of innovative deep learning methods for autonomous UAV-based railway maintenance system

The main purpose of this thesis is to automatically monitor railway lines and components and detect defects with high accuracy. The thesis aims to increase efficiency by minimizing labor and time loss with autonomous systems that replace traditional methods. One of the main motivations of the study is to detect foreign objects on rail lines and components. Such objects pose a significant risk to railway transportation safety, as they can negatively affect rail line safety. The method developed in the simulation environment enables the developing of more effective and reliable railway maintenance systems by adapting it to real-world conditions. Within the project's scope, flight simulations were carried out along the railway line using the Parrot Anafi4K UAV in the Gazebo simulation environment. In the studies to be carried out in the real-world environment, the images collected by the UAV will be processed with hybrid deep learning-based models to detect defects in railway components. The study aims to detect defects such as cracks and deformed fasteners in railway components with high accuracy by using deep learning algorithms such as YOLO, Unet, BiSeNetV2, and hybrid methods. In this way, railway maintenance operations can be carried out faster and less costly. In addition, an improved DNet segmentation model based on Unet has been developed, which outperforms the traditional Unet model. The findings of the developed methods were presented in detail by comparing them with the studies in the literature and successful and effective results were obtained.