面向无人艇视角的海上光学图像目标检测算法研究
2025-11-13王玉松纪延锯陈勤辉
中图分类号:E917 文献标志码:A DOI:10.3969/j. issn.1673-3819.2025.05.007
Abstract:During theapplication of unmanned surface vehicles (USVs)innatural sea areas,issuessuch as low-quality opticalimageimaging,localizedtargetfeaturesanddepth-directionaldistortioncausedbyside-viewingangles,lensdeforation duetosaltwaterdroplets,andthecoexistenceofextremelylargeandsmalltargetsocur.Thesephenomenaleadtoadecline intheperformanceofdeeplearning-basedimage targetdetectionalgorithms,resulting inhighratesofmisseddetectiosand 1alarms.Consequently,thiscauses thecolisionavoidancedecision-makingalgorithms todivergeortriggerfrequentemergencyalarms.To enhanceobjectdetection acuracyfor USVoptical images,,this paper proposes anoptimizedframework basedontheYOLOv8 model.First,theCopy-Pastealgorithmandunpaired image style transferalgorithmare introduced to generate multi-viewtargetimagesandlow-qualityimagescausedbyadverseenvironmentalconditions,addressing theimbalanceintheproportionoflow-qualityimagesinthedataset.Second,adedicateddetectionheadforsmalltargetsisadedto thedetectionnetwork,andthelossfunctionisoptimizedtoenhancethedetectioncapabilityforsmalltargetswhilepreserving theoriginal model’sperformanceonlargerobjects.Thenewlydevelopedobjectdetection model,trainedontheaugmented dataset,achieves an average precision of 96.2% on the test set.
Keywords:USV;YOLOv8;data augmentation;target detection
随着人工智能、传感器技术等飞速发展和深度应用,自主无人艇得以蓬勃发展,成为海洋开发与利用的新质平台。在复杂多变的自然海洋环境下,无人艇自主航行能力发展的关键因素是环境感知及处理能力。通常情况下,无人艇配置多种传感器在不同的感知域对无人艇周边进行联合感知测量,以弥补单个传感器的能力不足,如光学传感器、激光雷达、毫米波雷达、导航雷达等设备[1],其中,光学传感器具备模块化程度高、技术成熟,经济性好,无固有盲区,成像细节特征丰富等优点,是无人艇标配设备,在无人艇近区环境感知中扮演着至关重要的角色。
光学传感器对环境干扰尤为敏感,在无人艇应用场景中,在侧视视角下,目标呈现特征局部化和轮廓深度方向形变(见图1);雨、雪、雾、强光、逆光、弱光等光照条件影响,导致成像质量不高,波浪及浪花形成形状各异的亮块、阴影、光晕等影响目标检测效果(见图2);镜头水滴和盐渍污染导致图像局部透镜效应或椒盐噪声(见图3)等。上述因素使图像目标检测算法在无人艇应用场景下性能不稳定,自主任务过程中,频繁造成停船或紧迫局面,对无人艇自主执行任务的安全性和连续性造成影响。本文主要针对无人艇在自然海域中光学传感器成像质量不高、侧视视角图像的特点,开展目标检测算法研究,以提升目标检测效果。
图1同一目标在侧视视角下特征变化显……
