Analysis of Visual-Inertial Techniques for UAV Navigation in GPS-denied Urban Environments

This thesis investigates the autonomous navigation of Unmanned Aerial Vehicles (UAVs) in GPS-denied urban environments, focusing on a Visual Inertial Odometry (VIO) algorithm developed from scratch for vehicle pose estimation. In contexts where GPS signals are limited or not available, such as urban areas, VIO provides a promising solution by integrating visual information from a monocular camera with inertial data from an Inertial Measurement Unit (IMU). A monocular visual-inertial system acts as the minimal sensor suite for six degrees-of-freedom (DOF) metric state estimation, offering advantages in size, cost and simplicity of hardware setup. The main objective of this research is to design a monocular visual-inertial odometry algorithm using a loosely coupled approach that enables drones to navigate autonomously using only onboard sensors. A significant challenge in monocular Visual Odometry is scale ambiguity. This problem occurs because a three-dimensional scene is reduced to a two-dimensional image plane, losing depth information. Despite this limitation, monocular systems are still a cost-effective choice and are often favored over stereo vision, especially when the distance of the scene from the camera is much greater than the stereo baseline. This issue is addressed through the fusion of IMU and camera data using an Error State Kalman Filter (ESKF), which combines visual information with inertial measurements to resolve scale ambiguity and achieve reliable pose estimation. This work also includes a thorough review of key visual localization components, analyzing various design approaches to assess their advantages and limitations. As feature extraction algorithm, the SIFT method is proposed. A detailed examination of SIFT is also done to understand its structure, feature extraction methods and its characteristics in general. Additionally, a theoretical analysis of the essential matrix and its computation methods is performed, focusing on the five-point algorithm used in 2D-to-2D methods for estimating camera motion in visual odometry. A synthetic dataset is created using the integration of PX4, Gazebo and QGroundControl frameworks. PX4 provides the flight control software, including necessary drone dynamics and control algorithms. Gazebo is used to develop a realistic 3D simulation environment, featuring various elements like buildings, roads and obstacles that closely replicate real-world conditions. Incorporating these details, the tailored environment provides a valuable testing ground truth for analyzing the effectiveness of the proposed algorithms in urban navigation scenarios. QGroundControl acts as the ground control station, enabling mission planning and monitoring of the UAV during simulations. Therefore, this simulation setup allows for the design of various trajectories and the generation of useful data for analysis, as well as demonstrating the VIO algorithm's performance.