Study of SLAM State of the Art Techniques for UAVs Navigation in Critical Environments

In recent years, Autonomous Navigation for flying robots has become a global challenge. Unmanned Aerial Vehicles (UAVs) have a high potential in both military and civil applications such as aerial reconnaissance, surveillance, search and rescue, product deliveries, agriculture, or infrastructure inspections. Autonomous Navigation of UAVs exploits an onboard Inertial Measurement Unit (IMU) which consists of a three-axis accelerometer and a three-axis gyroscope providing the linear acceleration and the angular velocity of the robot. The problem is that IMU measurements suffer from noise and bias resulting in a drift on the pose estimation. Even if the drift is irrelevant, it will accumulate to a significant value over time. In outdoor navigation, this issue seems to be solved by fusing IMU measurements with data from an onboard GNSS/GPS. Unfortunately, in indoor navigation, there’s no possibility to use this technology for the state estimation of the UAV because of the GNSS/GPS signal which is usually degraded or totally not available. This means that Autonomous Navigation of UAVs in GPS-denied environments is still an open challenge. Over the last few years, the scientific community has focused attention on vision-based navigation thanks to the latest innovations in embedded hardware solutions, which allow us to have at the same time a high computational power and low weight to satisfy the payload constraints of aerial platforms. This Master Thesis fits into the context of the Leonardo Drone Contest, a three-year competition launched by Leonardo to motivate young researchers to improve Artificial Intelligence applied to UAVs, in which six Italian universities compete against each other. The purpose of this work is to study the SLAM State of Art Techniques for UAVs Navigation in Critical environments. Therefore it provides a comparison between different visual-inertial algorithms to assess which one is the best solution in terms of accuracy for navigation in GPS-denied environments. The analyzed algorithms are the highly optimized proprietary VI-SLAM algorithm of the Intel T265 Tracking Camera, a semi-direct monocular VIO algorithm, SVO, an optimization-based VIO algorithm, VINS-Fusion, and the VI-SLAM Systems, ORB SLAM 2 and its improvement, ORB SLAM 3. The above-mentioned algorithms are tested along a path in a small indoor environment full of features with artificial lighting using an INTEL T265 Tracking Camera. The algorithms’ performance is evaluated by the study of the localization error with respect to the ground-truth reference path. The results show that the best choice in terms of accuracy is VINS-Fusion. Nevertheless, its excellent performance requires a high level of computational resource usage. Consequently, using this algorithm onboard for state estimation needs a preliminary check on how many computational resources will remain for navigation, control, and other essential tasks.