基于混合模型的山地高速公路能见度等级评估

Assessment of Visibility Levels for Mountainous Highways Based on Hybrid Models

摘要:

贵州高速通车里程全国第五，但交通气象站布设稀缺，山地地形影响下低能见度天气频发，严重威胁道路交通安全。借助深度学习图像识别方法识别能见度等级，可辅助雾天交通管制快速评估。卷积神经网络（Convolutional Neural Network, CNN）受限于全局特征获取缺失，分类精度不高，本文兼顾局部与全局特征表达，构建融合CNN与Transformer的混合网络模型cTrans-Net，选取贵州典型山地高速监控视频图像参与模型训练和测试，实现影响交通的能见度0～4共五级分类。结果表明，cTrans-Net测试集总体准确率达89.17%，ROC曲线下面积为0.9822，优于多种主流深度学习模型；独立验证集的总体准确率为87.75%。在对交通有明显影响的较低能见度（＜500 m）的评估中，针对发生频次最高的2级（界于100～200 m）召回率最高为90.66%，在0（＞500 m）与4（≤50 m）两级低频样本上的召回率分别达92.31%和90.45%，cTrans-Net模型适应于样本不均衡的场景，实用性强。特征可视化结果显示，cTrans-Net聚焦交通标线及雾气等利于能见度评估区域。本研究尝试为智慧交通中的雾天识别提供技术方案。

Abstract:

Guizhou Province ranks the fifth in China in terms of expressway mileage, yet the deployment of traffic meteorological stations remains sparse. Frequent low-visibility weather events, influenced by mountainous terrain, pose severe threats to road safety. Deep learning-based image recognition methods for visibility level classification assist in rapid assessment for traffic management during foggy conditions. However, conventional neural networks (CNNs) suffer from limited classification accuracy due to insufficient global feature extraction. To address this issue, this study integrates both local and global feature representations by constructing a hybrid network model, cTrans-Net, which combines CNNs and Transformer architectures. Surveillance video images from typical mountainous expressway sections in Guizhou are selected for model training and testing to achieve precise five-level visibility classification (L0-L4) that impacts traffic safety. Experimental results demonstrate that cTrans-Net achieves an overall accuracy of 89.17%, with an area under the curve (AUC) of 0.9822, outperforming several mainstream deep learning models. The overall accuracy of the independent validation set remains 87.75%. In evaluating low-visibility conditions (<500 m), which significantly affect traffic, the model attains the highest recognition recall of 90.66% for the most frequently occurring L2 level (100-200 m). For the less frequent L0 (>500 m) and L4 (≤50 m) categories, the recognition recall reaches 92.31% and 90.45%, respectively, indicating the model’s robustness in handling imbalanced datasets. Feature visualisation reveals that cTrans-Net effectively focuses on key regions such as road markings and fog distribution, which are critical for visibility assessment. This study provides a technical solution for fog-related visibility recognition in intelligent transportation systems, offering practical value for real-world applications in mountainous expressway environments. The proposed cTrans-Net demonstrates strong adaptability to imbalanced data scenarios and exhibits superior performance in critical low-visibility conditions, making it a viable tool for enhancing traffic safety management under adverse weather conditions.