摘要
鱼眼镜头非线性光学畸变导致鱼眼图像行人检测算法精度低,且校正算法也无法完全克服鱼眼图像的边缘严重变形。针对上述问题,文中以Faster R-CNN架构为基础,建立了鱼眼图像校正光路模型。针对鱼眼图像畸变,提出一种基于微分方程的鱼眼图像校正模型,并提出一种改进算法用于鱼眼图像的行人检测。构建了ResNet 50融合特征金字塔网络结构,以增强网络的多尺度特征提取能力,提高网络对行人小目标的定位和识别能力;优化平滑L1损失函数解决大梯度难学样本与小梯度易学样本间的不平衡问题,提高训练效果。实验结果表明,文中算法与现有鱼眼图像行人检测算法相比,检测精度提高了39.68%。在边缘轻微畸变及小尺度行人的检测精度可以达到90%以上,有助于提高极端条件下鱼眼图像的行人检测性能。
The nonlinear optical distortion of the fisheye lens leads to low accuracy of pedestrian detection algorithms for fisheye images,and the correction algorithm fails to fully overcome the severe edge deformation of fisheye images.Therefore,a fisheye image correction optical path model is established based on the Faster R-CNN architecture.A fisheye image correction model based on differential equations is proposed to address the distortion of fisheye images,and an improved algorithm is proposed for pedestrian detection of fisheye images.A ResNet 50 fusion feature pyramid network structure is constructed to enhance the multi-scale feature extraction ability of the network and improve its localization and recognition ability for pedestrians(small objects).The smooth L1 loss function is optimized to eliminate the imbalance between difficult-to-learn samples with large gradients and easy-to-learn samples with small gradients,so as to improve training effectiveness.The experimental results show that the detection accuracy of the proposed algorithm is improved by 39.68%,and its detection accuracy of slight edge distortion and small-scale pedestrians can reach over 90%in comparison with the existing fisheye image pedestrian detection algorithms.Therefore,it is helpful to improve the pedestrian detection performance for fisheye images under extreme conditions.
作者
张瑶
刘发炳
黄国勇
钱俊兵
阮爱国
沈忠明
ZHANG Yao;LIU Fabing;HUANG Guoyong;QIAN Junbing;RUAN Aiguo;SHEN Zhongming(Faculty of Civil Aviation and Aeronautics,Kunming University of Science and Technology,Kunming 650500,China;CGNPC Yuxi Huaning Wind Power Co.,Ltd.,Yuxi 652800,China)
出处
《现代电子技术》
北大核心
2024年第15期40-46,共7页
Modern Electronics Technique
基金
国家自然科学基金项目(62363018)
教育部科技发展中心产学研创新基金资助课题(2021JQR023)
教育部产学研合作教育项目基金(220602518-231116)
航天科工深圳(集团)有限公司委托项目(KG-CX-FK-20230714002)。
