@article{discovery10197444,
         journal = {Ocean Engineering},
          volume = {312},
            note = {Copyright {\copyright} 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).},
       publisher = {Elsevier},
           title = {Reliable LiDAR-based ship detection and tracking for Autonomous Surface Vehicles in busy maritime environments},
          number = {Part-3},
            year = {2024},
           month = {November},
            issn = {0029-8018},
          author = {Xie, Yongchang and Nanlal, Cassandra and Liu, Yuanchang},
             url = {https://doi.org/10.1016/j.oceaneng.2024.119288},
        abstract = {Environmental perception is a crucial requirement of Autonomous Surface Vehicles (ASVs) if required to perform tasks safely in a dynamically complex operational environment. Most existing methods for ship detection rely on camera-based methods, which are sensitive to environmental conditions and cannot directly provide spatial location information related to detected targets. To overcome this limitation, we propose a LiDAR-based ship detection and tracking framework that can be applied to busy maritime environments. The proposed framework consists of two functional modules: a ship detection and multi-object tracking. For ship detection, a modularised network structure was adapted, allowing for ease of switching between different types of detection network to prioritise either detection accuracy, detection speed or a compromise of both, depending on the task requirements. A Kalman Filter-based multi-object tracking method is also implemented to compensate for any detections that may have been missed as a result of ship motions or occlusions, relying solely on the detection results. We also collected the first-ever real-world LiDAR dataset for maritime applications across the River Thames and marinas, including a range of ship types, with lengths ranging from 5 m up to 40 m, and different hull types. The datasets are organised in a similar manner to the KITTI datasets, which can be easily applied to the well-developed point cloud detection networks. Remarkably, our methods achieve an overall detection accuracy of 74.1\% in the collected datasets. The proposed framework and dataset make LiDAR-based environmental perception feasible for implementation in ASVs and support development in the autonomous maritime navigation field.},
        keywords = {LiDAR-based perception; Ship detection; 
Deep learning; 
Object tracking; 
Autonomous Surface Vehicle}
}