不同场景下的交通能见度估算模型的应用  

Analysis of the application of traffic visibility estimation models in different scenarios

作　　者：刘旗洋 乔枫雪[1,2,3] 陈博 宋智超 季仁杰 魏超时 LIU Qiyang;QIAO Fengxue;CHEN Bo;SONG Zhichao;JI Renjie;WEI Chaoshi(Key Laboratory of Geographic Information Science,Ministry of Education,East China Normal University,Shanghai 200241,China;School of Geographic Sciences,East China Normal University,Shanghai 200241,China;Institute of Eco-Chongming(IEC),East China Normal University,Shanghai 202162,China;The East China Regional Air Traffic Management Bureau under the Civil Aviation Administration of China(CAAC),Shanghai 200335,China)

机构地区：[1]华东师范大学地理信息科学教育部重点实验室,上海200241 [2]华东师范大学地理科学学院,上海200241 [3]华东师范大学崇明生态研究院,上海202162 [4]中国民用航空华东地区管理局,上海200335

出　　处：《大气科学学报》2022年第2期179-190,共12页Transactions of Atmospheric Sciences

基　　金：上海市自然科学基金资助项目(21ZR1420400);国家自然科学基金资助项目(41730646)。

摘　　要：能见度监测是交通出行安全的重要保障,尤其对机场和高速公路的大范围低能见度的监测和预警更为重要。在传统人工目测方法的基础上,以激光透射能见度仪为代表的仪器测量方法更为准确,但存在探测范围小、维护成本高、全覆盖耗资大的局限性。为了克服以上缺陷,使交通能见度的估计更为灵活、高效,本文基于机场气象站点观测数据、机场大雾以及高速公路低能见度图像,构建优化三种不同场景下的能见度估计模型,并探讨了不同模型的适用性。1)基于气象站点观测的能见度估计,运用相关系数矩阵和特征重要性分析筛选出相对湿度、温度、水平风速3个变量,并考虑昼夜分别构建三元三次多项式拟合模型,模型的决定系数(R^(2))可达0.9以上;2)基于机场大雾图像的能见度估计深度学习模型,利用尺度不变特征变换方法提取图像关键点的特征向量,输入全连接神经网络(fully connected neural network)模型,加快训练过程并提高模型的可解释性;3)基于高速公路图像的能见度估计的反演模型,根据暗通道先验理论和能见度测量基本方程,计算大气光亮度和透射率,并根据图像距离信息得到单目图像的能见度,该方法无须预置目标物和像机参数,也不需要训练样本。本研究考虑了基于气象观测的物理模型与基于图像特征的深度学习方法,建立分别适用于具有机场常规气象观测,以及机场大雾低信噪比图像或高速公路低能见度单目图像时对交通能见度的估计,有效降低了能见度监测对观测仪器的依赖性。Visibility is an important physical quantity that reflects the degree of atmospheric transparency,and is closely related to people’s daily life and traffic travel.In this study,in order to make the estimation of visibility more flexible and efficient,three visibility estimation models are constructed and improved for different scenarios,and the respective applicability,advantages and disadvantages of the different models are analyzed.First,the visibility estimation is performed based on meteorological station observations,using correlation coefficient matrix and feature importance analysis to filter out the three variables of relative humidity,temperature and horizontal wind speed,and both day and night are considered to build a ternary cubic polynomial fitting model,which improves the overall fitting ability.Second,the deep learning model of visibility performs estimation based on images,and the scale invariant feature change method is used to extract the feature vector of key points of images,as the training of fully connected neural network model.Next,as the training data of the fully connected neural network model,the computational cost is reduced and the stability of the model is improved.Third,the inverse model of visibility estimation based on height highway images,according to the dark channel a priori theory and basic equation of visibility measurement,the atmospheric luminosity and transmittance are calculated,and the visibility of the monocular images is obtained based on the image distance information.The method does not require pre-set target and camera parameters,nor does it require training samples.The three visibility estimation models can be adapted to different scenarios,and can reduce the dependence on observation equipment.

关 键 词：能见度估算 尺度不变特征变换 全连接神经网络 消光系数 暗通道先验 

分 类 号：U492.8[交通运输工程—交通运输规划与管理] P412.17[交通运输工程—道路与铁道工程]