基于多通路融合网络的高速公路雾天能见度等级识别

雾天是影响高速公路交通安全的重要因素。研究从监控图像进行高速公路雾天能见度的自动识别方法可以为交通管理部门的智能管理和决策提供技术支持。根据大气散射模型分析出与雾浓度相关的多个物理因素，提出了综合这些物理因素的多通路融合识别网络。该网络使用三个通路联合学习深度视觉特征、传输矩阵特征和场景深度特征，并设计注意力融合模块来自适应地融合这三类特征以进行能见度等级识别。同时构建了一个合成数据集和一个真实的高速公路场景数据集，用于网络参数学习和性能评估。实景数据集中的图像是从中国多条高速公路的监控视频中收集的。在这两个数据集上的实验表明，所提方法可以适应不同的监控拍摄场景，能够比现有方法更准确地识别能见度等级，有效提升了识别精度。

Recognition of highway visibility level in foggy weather using multi-stream deep fusion network

Foggy weather is an important factor affecting highway traffic safety. Research on the automatic recognition method of highway fog visibility from surveillance images can provide technical support for the intelligent management and decision-making of the traffic management department. This paper analyzed multiple physical factord related to fog density based on the atmospheric scattering model and proposed a multi-channel fusion network that integrates these physical factors. Specifically, the method jointly exploited three streams to learn deep visual feature, transmission matrix feature and scene depth feature, and designed an attention fusion module to adaptively fuse these three streams for the final visibility level recognition, which was very beneficial for improving the recognition accuracy. Meanwhile, this paper constructed a synthetic dataset and a real-scene dataset for network parameters learning and performance evaluation. The images in the real-scene dataset were collected from surveillance videos of multiple highways in China. Experiments on these two datasets show that this method can identify visibility level more accurately than existing methods.