Advancing Maritime Perception: PV-RCNN for 3D LiDAR Object Detection

Abstract The perception module of an Uncrewed Surface Vehicle (USV) is the basis for performing missions in real-world environments like harbours, rivers, seas, and water canals. Compared to manned systems, the advanced technology of USVs offers a wide range of applications in maritime infrastructures, such as inspections, transportation, and surveillance tasks or multi-agent combinations with other robotic systems like Autonomous Underwater Vehicles (AUVs). Nonetheless, maritime environments present certain challenges, e.g., disturbances like winds and waves with dynamic obstacles like ship traffic, buoys, and other moving objects. An USV needs to perceive and avoid these obstacles to ensure robust navigation and safe operation. Light Detection and Ranging (LiDAR)-based perception algorithms are widely used in robotic obstacle detection and avoidance due to their reliable and robust depth data measurements. The paper explores a deep-learning-based approach using PV-RCNN. The model takes raw LiDAR points as input and outputs classified boat detections. A LiDAR dataset was created with manually labeled boats and extended through augmentation techniques, such as rotation, scaling, and GT-sampling. The model was trained and evaluated with different hyperparameter settings, with the goal of improving autonomous navigation of USVs in maritime environments. Experiments showed that applying a moderate rotation of 10° during augmentation achieved optimal results in Recall and detection performance at IoU=0.5/0.7, leading to improved generalization and robustness compared to both no augmentation and higher rotation degrees. In contrast, longer training durations and extensive augmentation with high rotation angles led to low performance values. Altogether, the experiments demonstrate that adding more diverse data and optimizing dataset configuration, model architecture, and hyperparameters (e.g., higher batch size, more training epochs, additional object classes) can improve detection performance. These improvements contribute to more robust and reliable boat detection.