Research And Implementation Of Highway Weather Classification Based On Deep Learning

Weather conditions affect people’s production and life in all aspects.Especially in the field of transportation,bad weather conditions will not only reduce the efficiency of transportation,but also affect the road safety all the time.It is one of the main causes of traffic accidents on the highway.Therefore,the study of weather detection in highway scene is of great significance of highway management,traffic operation and so on.With the rapid development of computer performance and deep learning,image classification based on deep learning has made great progress,but there are still many problems and challenges about weather image classification based on highway scene.For example,there is a lack of public weather dataset for highway scenarios;The image background of highway is complex,and there are many interference information such as vehicles,vegetation,guardrails and pedestrians;weather information is widely distributed in the image,and there will be some correlation between the information in different regions.To solve the above problems,this thesis explores and studies the weather image classification task in highway scene of the perspectives of semantic segmentation,multi-task learning and attention mechanism.The main research work of this thesis is as follows:(1)In view of the lack of public weather image dataset of highway scene,based on road monitoring video and network video image resources,this thesis constructs a highway weather dataset FHWD(Five-class Highway Weather Dataset)with five weather types of sunny,cloudy,rainy,snowy and foggy,which includes various highway sections,shooting angles and background information.(2)Analyzing the image of highway scene,it is found that the key of weather classification lies in the sky and road area.Combining the advantages of feature pyramid network and dilated convolution,this thesis proposes a semantic segmentation structure DPUM(Dilated Pyramid Upsample Module)for sky and ground segmentation.MWNet(Multi-Task Weather Networks)is proposed,which takes the semantic segmentation of sky and ground as an auxiliary task to reduce the interference from irrelevant information on weather characteristics and enhance the weather characteristics of sky and ground areas.(3)In order to mine the association with different regions and channel features,a grouped attention module(GSAM)is proposed.The GSAM module divides the feature map into multiple groups of sub features,realizes the global spatial attention and channel attention extraction through the self-attention mechanism within the group,and models the relationship between different sub features of groups to highlight important sub features.Through experiments and class activation map,it is proved that the grouping attention module proposed to this thesis can improve the network performance.(4)Based on the analysis of the business requirements of managers and the weather classification algorithm proposed by this thesis,an intelligent highway weather detection system is designed and implemented.The system can monitor the weather of highway in real time and alarm the bad weather conditions.