摘要
为了实现近实时公里级的高速公路能见度监测,提出了一种基于多尺度融合网络的能见度估计方法。首先,从道路监控视频中提取道路场景图像,并对图像进行质量控制;分别采用引导滤波、光谱滤波、景深估计模型从路面场景图像中提取细节结构特征、光谱特征与场景深度特征;然后,构建多尺度融合网络自适应融合结构特征、光谱特征与场景深度特征,并从中提取能见度特征;最后在网络末端输出图像场景所对应的能见度等级。此外,为了训练与测试模型,构建了真实道路场景图像数据集,共包含18 000张标注图像。实验结果表明,多尺度融合网络可显著提升高速公路能见度估计的准确性,能见度等级分类准确率可达81.76%。
In order to realize near real-time kilometer-level highway visibility monitoring,in this paper,propose a new method of visibility estimation based on multi-scale fusion network was proposed in this study deep network architecture that directly estimates the visibility from highway surveillance images.Specifically,the road scene images were firstly extracted from road surveillance videos with quality control,then used the estimation models are applies for extracting multiple image detailed structure and spectral features,as well as scene depth features extraction methods to extract the detailed structural,spectral,and scene depth features from images.After that,the multi-scale fusion network is designed to extract and fuse important features for visibility estimation adaptively,and the visibility level corresponding to the image scene is output at the end of the network.Besides,constructed a real-scene dataset,which contains a total of 18 000 annotated images,for model learning and performance evaluation.The experimental results demonstrate that the multi-scale fusion network can significantly improve the accuracy of highway visibility estimation,and the accuracy of visibility classification can be reach promoted to 81.76% with this new method.
作者
黄亮
肖鹏飞
薛梅
张振东
孙家清
周雪城
HUANG Liang;XIAO Pengfei;XUE Mei;ZHANG Zhendong;SUN Jiaqing;ZHOU Xuecheng(Key Laboratory of Transportation Meteorology,China Meteorological Administration(LATM-CMA),Nanjing 210041,China;Jiangsu Meteorological Service Center,Nanjing 210041,China;School of Artificial Intelligence and Information Technology,Nanjing University of Chinese Medicine,Nanjing 210046,China)
出处
《气象科学》
北大核心
2022年第5期668-675,共8页
Journal of the Meteorological Sciences
基金
中国气象局创新发展专项(CXFZ2022J069)
江苏省气象局科研项目(KZ202105,KQ202110)。
关键词
能见度估计
多尺度融合网络
图像识别
visibility estimation
multi-scale fusion network
image recognition
