Highway Visibility Detection In Foggy Weather Based On Deep Convolutional Networ

Reduced visibility in foggy weather poses a serious threat to the efficiency and safe operation of highways,and comprehensive and efficient visibility detection is of great importance to highway traffic management.With the spread of highway surveillance equipment,the construction of automatic visibility detection methods based on surveillance images allows for intensive,wide-area detection while reducing costs.Accurately detecting visibility from images is challenging due to differences in imaging equipment and conditions,and the complex non-linear relationship between surveillance image features and visibility classes.Because of the excellent feature learning capability of deep convolutional networks,this paper aims to investigate image visibility level detection methods based on deep convolutional networks.This paper first constructs a real scene dataset,and two detection methods are proposed and applied in practice.Quality methods and specific results include:(1)A siamese network-based method for visibility detection with reference images.This paper proposes a detection method based on the siamese network.The method takes as input the image to be detected jointly with a fog-free reference image and detects visibility by discriminating the difference in features between the two.The detection network is designed with siamese feature extraction branches and an embedded attention module for extracting high-dimensional features from the input image pairs.The method is also based on the idea of multi-tasking,with the main branch comparing the differences between the extracted features for visibility classification and a parallel auxiliary branch that strengthens the differentiation learning capability of the network by measuring the similarity between the features.The experimental results show that the method has good detection performance,improves the discriminatory ability of similar categories,and reduces the impact of different camera views.(2)Single image visibility detection method based on multi-stream fusion network.To improve the applicability of the method in the real-world scene,this paper further explores the visibility detection method that does not rely on a fog-free reference image only a single foggy image.This paper first analyses several physical factors associated with fog concentration from an atmospheric scattering model and then proposes a multi-stream deep fusion network model that integrates these physical factors.Based on this idea,the method jointly exploits three streams to learn deep visual feature,transmission matrix,and scene depth feature and designs an attention fusion module to adaptively fuse these three streams for the final visibility level detection.The experimental results show that this method can be adapted to different surveillance shooting scenarios and improves the performance of detecting visibility for a single image.(3)Prototype system for automatic detection of visibility level in fog on the highway.Based on the proposed single image visibility detection method,an automatic visibility level detection prototype system is designed and developed in this paper.The system provides a user-friendly visual interactive interface that allows the user to select models,load weights,select images,and perform visibility level detection online,meeting the needs of applications in real scenarios.