文献类型：期刊文献

中文题名：基于相机-激光雷达融合的智能投喂船与养殖网箱相对定位方法

英文题名：A Relative Localization Method for Intelligent Feeding Vessels and Aquaculture Cages Based on Camera-LiDAR Fusion

作者：李洁雯[1];李荣辉[1,2,3];刘乔[1,2,3];李佳文[1,2,3];杜煜城[1]

机构：[1]广东海洋大学船舶与海运学院,广东湛江524088;[2]广东省南海海洋牧场智能装备重点实验室,广东湛江524088;[3]广东省船舶智能与安全工程技术研究中心,广东湛江524088

年份：2026

卷号：46

期号：2

起止页码：147

中文期刊名：广东海洋大学学报

外文期刊名：Journal of Guangdong Ocean University

收录：北大核心2023、、北大核心

基金：国家自然科学基金面上项目(52571405);广东省普通高校高端装备制造(智能机器人)重点领域专项(2023ZDZX3003);广东省研究生创新计划项目(040510132301);广东省教学质量工程项目(010203132501)。

语种：中文

中文关键词：智能船;养殖网箱;相对定位;多传感器融合;点云聚类;YOLOv11;海洋牧场

外文关键词：intelligent vessel;aquaculture cage;relative localization;multi-sensor fusion;point cloud clustering;YOLOv11;marine ranching

中文摘要：【目的】提出一种基于相机-激光雷达融合的相对定位方法,以实现海洋牧场智能船在动态环境下对漂移网箱自主投喂作业中的高精度相对定位。【方法】利用YOLOv11实时检测网箱上的参考标记并计算粗定位;进而引入视觉信息引导的自适应基于密度有噪声空间聚类(DBSCAN)的点云聚类机制,解决因点云密度变化导致的过分割与漏分割问题,从而精确提取候选目标在船舶坐标系下的位置信息;通过匈牙利算法实现跨模态数据关联,并基于置信度模型对关联结果进行自适应加权融合,从而动态优化船舶相对于网箱的相对位姿估计。【结果】湖面实船实验表明,在(15,25]m内,所提方法误检率和漏检率分别为2.7%和23.9%,较仅相机方法分别降低4.6%和14.9%,较仅激光雷达方法分别降低8.0%和7.7%,较加权融合方法分别降低2.2%和4.7%。同时,在(1,8]m内,所提方法平均距离误差为0.12 m,角度误差2.6°,较仅相机方法分别降低0.56 m和3.7°,较加权融合方法分别降低0.44 m和2.0°,且在(8,15]和(15,25]m区间内均保持稳定性能。【结论】所提轻量级相机-激光雷达融合相对定位方法能够针对目标距离变化进行自适应处理,提升多传感器的观测与融合效果,可为智能船自主投喂等近距离作业任务提供可靠的相对位姿感知支撑。

外文摘要：【Objective】This study aims to propose a relative positioning method based on camera-LiDAR fusion,so as to achieve high-precision relative localization for intelligent vessels during autonomous feeding operations of drifting fish cages in dynamic environments.【Method】YOLOv11 was employed to detect reference markers on the fish cage in real time,providing coarse localization estimates.A vision-guided adaptive density-based spatial clustering of applications with noise(DBSCAN)clustering algorithm was then introduced to address over-segmentation and under-segmentation issues caused by varying point cloud densities,thereby enabling accurate extraction of candidate target positions in the vessel coordinate frame.Cross-modal data association was performed using the Hungarian algorithm,followed by confidence-based adaptive weighted fusion to dynamically optimize the vessel-to-cage relative pose estimation.【Results】Lake-based field experiments demonstrate that,within the range of(15,25]m,the proposed method achieved false detection and missed detection rates of 2.7%and 23.9%,respectively.These figures represent reductions of 4.6%and 14.9%compared to the camera-only method,8.0%and 7.7%compared to the LiDAR-only method,and 2.2%and 4.7%compared to the weighted fusion method.Furthermore,within the range of(1,8]m,the proposed method exhibited an average distance error of only 0.12 m and an angular error of 2.6°,reducing of 0.56 m and 3.7°compared to the camera-only method,and 0.44 m and 2.0°compared to the weighted fusion method,with the stable performance across the(8,15]m and(15,25]m intervals.【Conclusion】The proposed lightweight camera-LiDAR fusion-based relative localization method can adaptively process according to target distance variations,effectively enhancing multi-sensor observation and fusion performance.It provides reliable relative pose perception support for closerange operational tasks such as autonomous feeding by intelligent vessels.