摘要
针对地铁场景下旋转小目标难以检测的问题,提出地铁场景下基于高分辨率网络和旋转目标框的旋转小目标检测方法。采用高分辨率网络和特征金字塔网络特征融合头部以增强特征提取能力,并提出旋转高斯核和旋转目标检测框的策略以适应任何方向的目标检测。实验结果表明这些改进使得模型检测准确率在旋转小目标、倾斜的细长目标检测上远远高于应用水平检测目标框的模型。该方法在地铁场景下能准确地检测水平目标的同时也能准确检测旋转小目标以及细长倾斜目标。
Aiming at the problem that rotating small objects are difficult to detect in subway scenes,this paper proposes a rotating small object detection method based on high-resolution networks and rotating target frames in subway scenes.The high-resolution network and FPN feature fusion head were used to enhance the feature extraction ability,and the strategy of rotating Gaussian kernel and rotating target detection frame was proposed to adapt to target detection in any direction.The experimental results show that these improvements make the model detection accuracy in the detection of rotating small targets and oblique slender targets much higher than that of applying horizontal detection target boxes.The method can accurately detect horizontal targets in the subway scene,but also can accurately detect rotating small targets and slender inclined targets.
作者
罗显光
李林
李振
陈洁
杨颖
黄志华
肖黎亚
Luo Xianguang;Li Lin;Li Zhen;Chen Jie;Yang Ying;Huang Zhihua;Xiao Liya(CRRC Zhuzhou Electric Locomotive Co.,Ltd.,Zhuzhou 412000,Hunan,China;Zhuzhou Guochuang Track Technology Co.,Ltd.,Zhuzhou 412000,Hunan,China)
出处
《计算机应用与软件》
北大核心
2025年第11期158-165,182,共9页
Computer Applications and Software
基金
国家重点研发计划项目(2019YFB1405404)
湖南省高新技术产业科技创新引领计划项目(2021GK4014)
株洲市科技计划项目2021关键技术攻关及成果转化项目(2021-003)
城轨车辆智能运维平台开发及应用项目(2020GKC4005)。
关键词
地铁场景
高分辨率网络
旋转小目标检测
旋转高斯核
Subway scene
High-resolution network
Rotating small target detection
Rotating Gaussian kernel
作者简介
罗显光,教授级高工,主研领域:图像识别、大数据智能运维系统的研究及应用;李林,高工;李振,高工;陈洁,高工;杨颖,高工;黄志华,高工;肖黎亚,高工。
