Research On Deep Network Model For Highway Visibility Detection In Foggy Weathe

Visibility is greatly reduced in thick fog and fog weather,which seriously affecting the traffic safety.With the popularity of highway video surveillance system,the management department can obtain monitoring images along the highway in time.The construction of image-based visibility detection method in foggy days can realize a wide range of visibility detection while utilizing the existing video surveillance system,and can provide timely warnings for transportation management departments and drivers.It is of great significance to improve the driving safety of highway in foggy days.However,due to the differences in camera equipment and imaging conditions,accurate detection of visibility from a single foggy image is still a challenging problem.In view of the excellent learning ability of deep network,this dissertation deeply studies the deep network model of highway foggy visibility detection and collects several foggy images on highways to build a data set,which includes:(1)Cross-scale feature context fusion network for highway visibility detection.In this paper,a cross-scale feature context fusion network is proposed,including two parts: dualbranch feature extraction backbone network and cross-scale feature context fusion.Specifically,on the one hand,the multi-layer convolution and pooling operation is used in the backbone network to extract the road global feature map with high-level semantic information,on the other hand,the multi-scale feature fusion strategy is used in the parallel branch to extract the multi-scale detail features.And designed a cross-scale feature adaptive fusion module to calculate the correlation between scales,and the context information is fused between multiscale detail features and the road global feature for final classification.Experiments show that the network can effectively learn the global and local features of images,and effectively improve the visibility detection accuracy.(2)Highway visibility detection in foggy weather of united lane lines geometric prior.In this paper,a dual-branch fusion network driven by geometric and visual features is proposed to realize visibility detection from a single fog image.Specifically,the proposed network consists of two branches,namely,the geometric feature branch and the visual feature extraction branch,and the final classification is carried out by the fusion classification module.In the geometric feature branch,the lane line detection model is designed to extract the lane line information in the expressway.The visual feature branch aims to directly learn the depth features of foggy monitoring images.In the fusion classification module,the lane length estimation module is designed to further extract the geometric length information of the lane lines,and the road attention module focuses on extracting the depth features of the road area.Finally,the two types of features are combined to detect the visibility level through the full connection layer.Experimental results show that this method can make full use of the lane information,improve the visibility detection accuracy,and has superior detection performance.