Research And Implementation Of Highway Bad Weather Detection Based On Deep Learning

In the context of the rapid development of highway construction in China,people are paying great attention to traffic safety issues,and bad weather conditions are an important factor affecting highway traffic safety.Therefore,it is of great significance to study the bad weather detection of highways for traffic safety.The traditional weather detection algorithm has great limitations,and it is difficult to meet the requirements of real-time monitoring of highway weather conditions.With the rapid development of deep learning technology,the highway traffic system is gradually developing in an intelligent direction,although the bad weather detection algorithm based on deep learning has achieved good results in performance,but there are problems of high computational complexity and large model capacity.How to design an efficient and lightweight highway bad weather detection algorithm has become an urgent problem to be solved.Based on the method of deep learning,this thesis studies the highway bad weather detection algorithm from multiple angles of network structure,convolution method,attention mechanism and image feature,and proposes a lightweight and accurate bad weather detection algorithm.The main work of this thesis is as follows:This thesis builds a highway weather dataset(Highway Weather Dataset,HWD)for various sources such as highway surveillance video,You Tube car video,Google pictures,etc.,covering various highway scenes,different camera angles,and various severity of different weather conditions.From the research of network structure and convolution method,the Mobile Net V2 network structure is optimized.Combined with the asymmetric convolution structure,an asymmetric convolution lightweight network structure ACMNet(Asymmetric Convolution Mobile Net)is proposed,which reduces the amount of parameters and improves the network performance.From the research of the attention mechanism,further optimize and improve on the basis of ACMNet,A group attention convolution network structure GACMNet(Group-attention Asymmetric Convolution Mobile Net)is proposed.The SGE attention module is optimized,and the maxpooling is selected as the local channel feature extraction method.By using Grad-CAM convolution to visually analyze the characteristics of network learning,the effectiveness of the attention mechanism for performance improvement is proved.From the research of image features,combine dark channel features with RGB features to form RGB-Dark four-channel features and send them to the GACMNet network for training,to improve the recognition effect of the network on foggy categories,By analyzing the comparison experiments of different fusion strategies and fusion functions,the RGB model and the RGB-Dark model were probabilistic fused by Max fusion function to balance the negative impact of the RGB-Dark model on the rainy and snowy categories,which further improves the performance of the highway bad weather detection algorithm.