无人水面艇上基于激光雷达的海洋物体检测

Marine object detection plays a crucial role in various applications such as collision avoidance and autonomous navigation in maritime environments. While most existing datasets focus on 2D object detection, this research introduces a novel 3D object detection approach that relies exclusively on LiDAR (Light Detection And Ranging) data, specifically tailored for small Unmanned Surface Vehicles (USVs), where energy efficiency and computational constraints are key challenges. This study contributes a new point cloud dataset collected from a 2-meter autonomous USV and augmented through a hardware-in-the-loop simulation environment. The PointPillars network, chosen for its efficiency in processing LiDAR data, was trained and evaluated in this maritime context. A comparative analysis was also conducted between the proposed LiDAR-only method and a multimodal (LiDAR-camera) approach. The core innovation of this work is a step for late fusion strategy, where object detection is performed independently across sensors before integration. This results in a significantly less resource-intensive solution compared to early fusion methods. Consequently, the LiDAR-only approach highly suitable for deployment on compact, low-power autonomous surface drones, marking a step forward in practical and scalable marine perception systems.