Uncertainty-Aware Maritime Point Cloud Detector (U-MPCD) for Autonomous Surface Vehicles

Autonomous surface vehicles (ASVs) operating in busy and constrained maritime environments (e.g., inland waterways, harbors, ports, and marinas) require robust perception modules for real-time boat detection, with LiDAR serving as one of the practical sensors for environmental perception. However, these environments present challenges, such as large variations in boat sizes, sparse point cloud data at longer distances, and occlusions from the restricted field of view of onboard LiDAR and surrounding obstacles, leading to high predictive uncertainty. Small boats rely on local features (e.g., fine-grained geometric details), while large boats require global features (e.g., overall shape and structural continuity) for accurate detection. To address these challenges, we propose the maritime point cloud detector (MPCD), which integrates an attention-based point feature net for pillar-level local feature extraction and a hybrid 2-D backbone combining multiscale MobileViT with a 2-D convolutional neural network for enhanced global feature learning, achieving a 12.8% improvement in detection accuracy over the baseline. To further enhance reliability, we extend MPCD with the multi-input multi-output method, forming uncertainty-aware MPCD (U-MPCD). U-MPCD estimates both epistemic and aleatoric uncertainties, improves detection accuracy by 2%, and maintains an inference speed of 15 Hz, providing critical insights into prediction confidence for safer ASV navigation. Our model was tested on real-world data sets collected under normal hydrographic survey conditions (6 h per day over four days, covering about 11.4 km) along the River Thames in central London, which features high maritime traffic and diverse boat types and sizes.