基于多特征聚合的水面遮挡目标检测算法  

Multi feature fusion-based water occlusion object detection algorithm

作　　者：冯辉[1,2] 蒋成鑫 徐海祥[1,2] 谢磊[3] FENG Hui;JIANG Chengxin;XU Haixiang;XIE Lei(Key Laboratory of High Performance Ship Technology of Ministry of Education,Wuhan University of Technology,Wuhan 430063,China;School of Naval Architecture,Ocean and Energy Power Engineering,Wuhan University of Technology,Wuhan 430063,China;Intelligent Transportation Systems Research Center,Wuhan University of Technology,Wuhan 430063,China)

机构地区：[1]武汉理工大学高性能船舶技术教育部重点实验室,湖北武汉430063 [2]武汉理工大学船海与能源动力工程学院,湖北武汉430063 [3]武汉理工大学智能交通系统研究中心,湖北武汉430063

出　　处：《华中科技大学学报（自然科学版）》2024年第4期76-81,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家重点研发计划资助项目(2019YFB1600600,2019YFB1600604);国家自然科学基金资助项目(51979210,51879210);中央高校基本科研业务费专项资金资助项目(2019III040,2019III132CG)。

摘　　要：针对智能船舶在内河航道航行时经常出现的船舶互相遮挡而影响目标检测精度的问题,提出了一种基于多特征聚合的水面遮挡目标检测算法.首先,在骨干网络设置多尺度感受野特征融合结构,融合被遮挡船舶可视区域与周围环境特征;其次,在骨干网络及网络的特征拼接部分添加混合注意力机制,增强网络的长程依赖性,聚合船首和船尾的特征;然后,设计了数据重采样策略,在训练过程中根据船舶类别的数量自适应地调整样本采样频率,缓解数据集中船舶数量的严重不均匀;最后对算法进行验证。结果表明:算法通过聚合被遮挡船舶可视区域与周围环境等多尺度特征,聚合船首、船尾长程特征,相较于原算法精度提升达到了3.3%,有效提升了视觉遮挡状态下水面目标的检测精度.Aiming at the problem that the object detection accuracy was affected by the mutual occlusion of ships that often occured when intelligent ships navigated in inland waterways,a water surface occlusion object detection algorithm based on multi-feature aggregation was proposed.First,a multi-scale sensory field feature fusion structure was set up in the backbone network to fuse the visible area of the occluded ship with the surrounding environment features.Second,a hybrid attention mechanism was added to the backbone network and the feature splicing part of the network to enhance the long-range dependence of the network,and to aggregate the features of the ship's bow and stern.Then,a data resampling strategy was designed to adaptively adjust the sample frequency according to the number of ship categories during the training process to alleviate the serious unevenness of the number of ships in the dataset.Finally,the algorithm was validated.Results show that the algorithm can effectively improve the detection accuracy of surface targets under visual occlusion by aggregating multi-scale features such as the visible area of the occluded ship and the surrounding environment,and by aggregating the long-range features of the bow and the stern of the ship,with an accuracy increase of 3.3%compared with the original algorithm.

关 键 词：智能船舶 遮挡检测 多尺度特征融合 混合注意力机制 数据重采样 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]