| Fog density levels can generally be divided by visibility,which is usually monitored by visibility instruments.However,the actual situation is that the visibility instruments are set at a large interval,and it is difficult to effectively monitor the sudden formation of agglomerate fog,which is a severe weather phenomenon that is extremely harmful on the expressway,about 300 to 800 meters in diameter.At present,there are surveillance cameras every 1-1.5km on the expressway.And estimating visibility by analyzing image content is one of the most effective and economical means to detect and warn the agglomerate fog.Therefore,this paper uses surveillance video on the expressway for analysis,and proposes two visibility estimation methods based on deep learning,using deep neural network models and lightweight network models,respectively,to extract lane line features and fog features,estimate road visibility,and quickly detect fog density level.There are continuous lane lines on expressways,the visible extension distance of lane lines in fog can roughly reflect the visibility of the road at that location.Therefore,this paper designs a deep neural network model to extract lane line features,and a visibility detection network estimates the road visibility based on the visible length of the lane line.This paper designs a U-shaped lane segmentation network based on the deep neural network VGG16.Then the model reduces the dimension of the feature map and inputs it to the visibility detection network to get the height of the lane in the image.Then this paper proves that there is a clear non-linear relationship between the value and visibility.The visibility can be accurately calculated by setting several parameters.Experiments show that this method can predict visibility from video images,especially for the detection of low visibility.In the case of high-visibility,the actual distance represented by one pixel increases at the end of the lane due to the camera shooting angle which causes the error of the method to become larger.This method maintains high accuracy in nine expressway sections.In order to improve the calculation efficiency and the detection accuracy of high visibility,this paper extracts lane line features and fog features based on the lightweight network model and designs a feature fusion network to predict visibility according to the lane line length,fog range and concentration.In view of the huge parameters of VGG16 and low computational efficiency,this paper uses the lightweight network model Bi Se Net V2 to replace the deep network as the main structure of the fog detection network,and to construct the lane segmentation branch.Aiming at the problem that the lane feature has a large error in predicting the visibility when the visibility is high,this paper uses the fog area feature and the image gradient map to assist the training.In view of the large-scale characteristics of the agglomerate fog,this paper improves the semantic branch of Bi Se Net V2 by embedding the feature refinement module and the feature fusion module,which can effectively retain global context information,and builds the dense fog area segmentation branch based on the improved Bi Se Net V2 network.Through two segmentation branches,the model obtains two kind of binary segmentation maps.The two segmentation maps are respectively fused with the gradient map in the feature fusion network.Then the fusion features are used to predict the visibility.Experiments show that the method effectively reduces the visibility detection time and improves the detection accuracy of high visibility.It basically meets the real-time requirements of fog detection and early warning. |
