COMPARATIVE ANALYSIS OF 2D AND 3D LIDAR-BASED OBJECT DETECTION FOR MARITIME NAVIGATION

Abstract

Maritime navigation safety relies on the timely and accurate detection of obstacles such as vessels, buoys, and floating debris. Conventional sensing technologies, including radar and optical cameras, are limited by low resolution, reduced performance in adverse weather, and susceptibility to sea clutter. Light Detection and Ranging (LiDAR), with its high-resolution three-dimensional sensing capability, offers a promising alternative; however, its application in maritime environments remains relatively underexplored compared to autonomous road vehicles. This study presents a LiDAR-based object detection framework for maritime navigation and evaluates the performance of two-dimensional (2D) and three-dimensional (3D) detection approaches using a calibration/projection-based LiDAR-camera fusion scheme, in which 2D detections are associated with LiDAR points to obtain 3D object localisation. A Velodyne mechanical LiDAR and a ZED stereo camera were mounted on a marine vessel operating in the waters of Klang, Malaysia. Training data were obtained from the Singapore Maritime Dataset, the SeaShips Dataset, and a custom-collected maritime dataset annotated using Roboflow for 2D images and labelCloud for 3D point clouds. A YOLOv8-based model was developed for 2D object detection, while a parallel pipeline generated point clouds from depth imagery for 3D bounding-box estimation. Performance was evaluated using precision, recall, F1-score, Intersection-over-Union (IoU), and, for the 2D pipeline, mean Average Precision (mAP50 and mAP50-95). The 2D detection model achieved strong performance, with a precision of 0.94, recall between 0.84 and 0.86, and an F1-score of approximately 0.89. In contrast, the 3D detection pipeline achieved precision, recall, and F1-score of approximately 0.25 each, a gap attributed jointly to the limited availability of annotated 3D maritime data and to the geometric, non-learned nature of the 3D bounding-box generation step, which takes the axis-aligned extent of the associated LiDAR/depth point cloud without a mechanism to reject noise or clutter points. Comparative analysis also showed that the Velodyne LiDAR produced denser and more accurate point clouds than the ZED stereo camera, particularly for distant objects where stereo depth estimation degraded. The proposed framework demonstrates the effectiveness of LiDAR-camera fusion for maritime 2D object detection while highlighting current challenges in 3D maritime perception. It provides an end-to-end baseline for future sensor-fusion research supporting safer autonomous and assisted maritime navigation.