Sensor fusion and visual-based localization for autonomous systems

One of the essential features for autonomous mobile robot applications that make their own decisions and perform tasks like humans is the necessity of delegating humanoid tasks to robots. An autonomous robot should understand the geometric structure of its environment, accurately localize itself, and create a motion trajectory guiding it to the specified point using this information. Sensor fusion obtained by combining different sensors plays a significant role in mobile robot localization, especially in cases where a single sensor is insufficient. In recent years, advancements in processor speed have led to increased focus on Visual Odometry (VO) methods using low-cost monocular cameras. Additionally, Visual-Inertial Odometry (VIO) solutions, incorporating low-cost Inertial Measurement Unit (IMU) sensors alongside cameras, have been preferred to contribute to localization. Traditional geometric-based solutions often struggle to accurately represent complex environments and face difficulties in obtaining reliable results. Therefore, Artificial Intelligence-based solutions are increasingly replacing traditional methods due to their adaptability to different environments. In light of the information mentioned above, this thesis proposes two different applications for developing autonomous vehicles in cases where a single sensor is insufficient. The first application presents a deep learning-based hybrid architecture using visual and IMU information to predict the position of a UAV (Unmanned Aerial Vehicle) moving indoors. The second application implements an artificial intelligence-based VIO (Visual-Inertial Odometry) application successfully predicting the position of an autonomous vehicle by processing consecutive camera images, using a different fusion technique. Both applications offer innovative methods for the localization of autonomous vehicles, demonstrating a tendency to outperform previous studies. Furthermore, it has been determined that the implemented applications are capable of operating in real-time systems.