Sensors Selection for Obstacle Detection: Sensor Fusion and YOLOv4 for an Autonomous Surface Vehicle in Venice Lagoon

Developments in the field of autonomous navigation have led recent studies to focus on addressing the challenges of maritime transportation in cities with rivers and canals. In this context, the construction of an Autonomous Surface Vehicle (ASV) capable of navigating through complex environments like the canals of the Venice lagoon presents a significant challenge. This thesis aims to elucidate key aspects of the perception part of this project. Initially, a comprehensive analysis of potential sensitization techniques for autonomous vessels is conducted to understand the appropriate sensor architecture. Achieving optimal environmental perception places great importance on the selection of sensors such as LiDAR and stereo cameras. On the other hand, choosing the right sensor fusion approach in the context of the selected collision-avoidance system is challenging. In the literature, various classifications of sensor fusion algorithms exist, and many versions, ranging from traditional to innovative. Finally, an object detection algorithm based on YOLOv4 has been developed using a dataset of images collected in the Venetian lagoon through Google Earth Pro. The detector distinguishes three different classes of typical Venetian boats giving in output parameters, as bounding boxes coordinates, reliability scores and probability with respect to the class to which it belongs, which are essential to perform an optimal obstacle avoidance.