Efficient Deep Visual–Inertial Odometry for robot localization

Autonomous navigation represents a cornerstone capability across a broad spectrum of robotic applications, such as self-driving vehicles and ground-based mobile robots. The precise estimation of the system’s motion over time is a process commonly known as odometry and it is central to achieving such autonomy. Over the years, numerous techniques have been developed to address this task, each offering distinct advantages and limitations depending on the sensors exploited and methodologies employed. Visual Odometry (VO) estimates the pose of a robot using images acquired from single or multiple cameras attached to the robot and has become one of the most robust techniques for vehicle localization. It extracts motion-related information from the apparent displacement of visual features across frames. This can be achieved using geometric relationships between camera poses and 3D scene points. Despite their effectiveness, purely visual approaches have drawbacks such as motion blur, change in lighting or insufficient visual elements. In recent years, Deep Learning has emerged as a powerful alternative to traditional visual odometry techniques. Data-driven models reduce the dependence on hand-engineered features and geometric models, learning robust representations directly from raw sensor data. This thesis focuses on the development of a hybrid deep learning-based visual-inertial odometry system, aiming to leverage the representational power of convolutional neural networks and recurrent architectures, while fusing this information with inertial data by means of a direct Extended Kalman Filter (EKF). The objective is to design and implement a system capable of pose estimation that does not require high performance hardware components. The key to achieving this goal, is the trade-off brought by the hybrid architecture, that limits the use of the deep neural networks to feature detection and feature matching. Following the development process of the system, the Deep VO algorithm alone was tested first, followed by the tests for the entire VIO pipeline. Moreover, for the sake of comparison, the Deep VIO technique under test was compared against another VIO algorithm, developed ad hoc for this aim, that doesn't rely on neural networks. In order to quantitatively assess the accuracy of the estimated trajectory with respect to the ground truth, several standard metrics are considered: absolute trajectory error, relative pose error, path length, final drift, Hausdorff distance and scale factor.