团雾是一种能见度低又瞬时变化快的气象现象，通常具有区域性特征，形成机制较为复杂造成其预测难度较高。团雾多发生于湿度较高的天气，路面湿滑，能见度又骤然降低，极易引发交通事故甚至是二次事故。因此，对高速公路上的团雾进行检测以辅助道路交通管理和安全预警具有重要的实践意义，本文开展了基于深度学习的智慧高速场景中团雾检测方法的研究。论文主要研究工作包括：1. 现在的雾天数据主要存在于用于自动驾驶研究的交通场景数据集中，数据量比较少，更没有专门的团雾数据集。针对雾天数据稀缺这一问题，本文利用贵州道坦坦科技股份有限公司提供的高速路段摄像头录制的视频数据，做取帧处理后，剔除模糊及拍摄角度过低或过高的异常图像数据，将图像按照无雾、轻雾、大雾、浓雾和强浓雾 5 个等级进行标注，制作了图片规模为 2477张的高速公路雾天场景数据集。针对没有团雾数据集的问题，本文使用 Carla仿真平台仿真生成了公路上的团雾，录制了不同时段的含不同浓度雾的交通场景视频。对视频帧进行提取后，将图像雾浓度划分为 5 个等级，然后标注创建了由仿真图像构成的团雾数据集，它由 1000 组时序图像序列构成，每一个图像序列由 10 张时间间隔为 3min 的按时序排列的图像组成。数据集用于支撑后续团雾检测方法研究。2. 针对雾浓度检测这一团雾检测核心问题，本文利用卷积神经网络对雾天公路视频图像进行雾浓度级别分类。使用 Resnet50 和 VGG16 卷积网络混合，再增加自适应模块，以此构建网络进行样本学习，这样可以避免过拟合。同时因为自适应结构的加入，实验验证模型对雾图像等级识别检测的收敛速度和识别精度都得到有效提高，针对 5 个级别能见度的雾等级识别准确率均在 0.8以上。3. 与以往仅通过能见度低这一特征甄别团雾相比，本文依托的 Carla 仿真数据集具有时序图像序列数据，结合图像序列中不同雾浓度图片表征的雾能见度短时间内发生变化这一团雾特征，论文进行了团雾检测。考虑未来工程应用，论文还设计了相应的综合考虑团雾时间特征及空间特征的高速公路团雾检测实施方案。
Agglomerate fog is a meteorological phenomenon characterized by low visibility and rapid intermittent changes. It usually exhibits regional features and its formation mechanism is complex, resulting in a high level of difficulty in prediction. Agglomerate fog may have a significant negative impact on highway traffic. This article presents a study on a deep learning-based detection method for agglomerate fog in intelligent highway scenarios.The main research work of this paper includes:1. From video data recorded from 35 highway cameras provided by Guizhou Daotantan Technology Co., Ltd., this paper extracted frames using ffmpeg and removed abnormal images with blurriness or excessively low or high shooting angles. Theremaining images were separated into five categories and labeled by labelme : no fog, light fog, heavy fog, dense fog, and strong dense fog. Finally we get a highway foggy scene dataset including 2477 images. To address the issue of limited agglomerate fog data, this study simulate agglomerate fog on roads by the Carla simulation platform and recorded 31 traffic scene videos with different fog densities. Frameswere extracted using ffmpeg, and the images were divided into five fog density levels and labeled by labelme. We create a agglomerate fog dataset consisting of 1000 groups of image sequences with ten images per sequence arranged in time intervalsof 3 minutes. This dataset is used to support subsequent research on agglomerate fog detection methods.2. To address the core issue of agglomerate fog detection, which is fog density classification, a convolutional neural network (CNN) was utilized to classify fog density levels in highway video images captured in foggy weather. A hybrid CNN consisting of Resnet50 and VGG16 was constructed, along with an adaptive moduleto avoid overfitting during sample learning. The addition of the adaptive structure resulted in improved convergence speed and recognition accuracy for fog image level recognition and detection. The accuracy of fog level recognition for the five visibility levels was above 0.8.3. Compared with previous methods that only detect agglomerate fog based on the low visibility feature, this paper utilized the Carla simulation dataset, which contains time-series image sequence data, to detect agglomerate fog based on the characteristic that visibility in the image sequence changes rapidly with different fogdensities. The paper also designed a comprehensive agglomerate fog detection implementation plan for highways that involves both temporal and spatial features of agglomerate fog, taking into account future engineering applications.
