深度学习在危险天气自动识别中的应用研究

Research on automatic identification method of traffic dangerous weather based on deep learning

基于机器视觉的危险天气自动识别技术近年来已成为研究热点,但模型识别准确率不高和模型不够轻量化是该项技术面临的主要问题。针对上述问题,提出了一种利用CycleGAN网络自动扩展危险天气数据集的方法,有效解决了数据集数据量不足、数据类型不平衡的问题。同时,还提出了一种三通道融合卷积神经网络(3-Channel Convolutional Neural Network,3C-CNN),该网络主干分支采用迁移学习的技术方案,并利用多分支结构提取并融合天气图像中的整体与局部特征。结果表明,利用CycleGAN网络扩充的WeatherDataset-6Plus数据集能够有效改善深度学习模型的训练性能,3C-CNN模型的6类天气现象综合识别准确率达到了98.99%,识别速度达到220帧/s。该方法在保证准确率的同时实现了模型的轻量化,有利于其在嵌入式设备中部署。

Recently,automatic recognition of dangerous weather based on machine vision has been gradually applied in automatic driving.However,the recognition accuracy of the popular models is not satisfactory,and the models are far from lightweight,which are the main challenges of this technology.Aiming at these issues,an elaborated method is presented in this paper to automatically extend dangerous weather data sets by using a CycleGAN network.The CycleGAN model is trained with an Adam optimization algorithm.As a result,the existing four types of weather data sets of sunny,rainy,foggy,and snow are expanded into six types of weather data sets:sunny,cloudy,rainy,foggy,snow,and dust.Thus,the problems of insufficient data and unbalanced data types in data sets can be solved effectively.To better extract features of weather images,a 3-Channel Convolutional Neural Network(3C-CNN)is constructed in this thesis,and a migration learning schema is adopted on its backbone.A multi-branch structure is designed to extract the sky features,ground features,and local features respectively in the weather images and achieve three feature graphs of three channels.Then,the Concatenate function is used to fuse the three feature graphs according to channel dimension and the classification results are finally obtained.The restrictions of insufficient data sets,the low efficiency of the classical deep learning networks in fusing the sky and ground features when extracting image features,and the lack of lightweight that is more suitable for deployment in onboard embedded devices are broken through in our proposed model.The experimental results show that the training performance of the deep learning model is significantly improved by our WeatherDataset-6Plus dataset,which is expanded by the CycleGAN network.A 98.99%comprehensive recognition accuracy of the 3C-CNN model for six types of weather phenomena is obtained,and the recognition speed Frames Per Second(FPS)is promoted to 220 as well.Also,an elegant balance is achieved between the recognition accuracy and model parameters.