Studies on Perceptive Systems for Maritime Autonomous Surface Ships

The development of Maritime Autonomous Surface Ships (MASS) represents a pivotal advancement in maritime safety and operational efficiency, particularly by mitigating accidents stemming from human error and reduced visibility. While the automotive and aerospace industries have made significant strides in autonomous perception, maritime situational awareness still relies primarily on traditional, non-cooperative sensors such as radar and the Automatic Identification System (AIS), which face substantial limitations in cluttered coastal and high-traffic settings. This dissertation tackles these gaps by adapting and validating advanced sensing technologies—such as LiDAR, thermal imaging, and RGB cameras—to the distinctive challenges of the marine domain. The core aim is to design, implement, and experimentally validate a multi-modal perception and communication framework that delivers high-level situational awareness for autonomous vessels. The study examines sensor fusion techniques, mitigates the shortage of maritime datasets, investigates AI-based target detection and classification, and assesses the transferability of perception algorithms from other domains to maritime applications. A structured experimental approach was adopted, advancing from simulation-based evaluations to controlled field trials and demanding sea tests. A modular sensing architecture was engineered, integrating LiDAR, thermal, and RGB sensors with high-precision INS/GNSS state estimation. To facilitate algorithm validation, two custom datasets were created: a LiDAR dataset comprising 11,680 scans (including 920 manually labelled samples) and a multi-modal "human-in-water" dataset comprising 6,000 samples, augmented with physical and virtual data to replicate environmental perturbations. Quantitative assessments on these datasets confirm the efficacy of the proposed pipelines. System-level validations on the SWAMP autonomous platform in Lake Nemi and the Gulf of Naples affirmed the seamless integration of perception and state estimation, supporting effective autonomous collision avoidance and robust multi-node communication. Multi-modal fusion was crucial in overcoming sensor-specific drawbacks, including LiDAR sparsity and thermal noise. This research contributes validated algorithms, novel datasets, and actionable guidelines for cross-domain adaptation of perception technologies to maritime contexts, laying a solid groundwork for future advancements in autonomous surface vessels.