Research And System Implementation Of Highway Vehicle Fine-Grained Classification Based On Deep Learning

With the increase in the number of vehicles on the highway,traffic management departments pay more and more attention to how to effectively classify and manage vehicles.At the same time,as the basic work of vehicle detection and traffic flow statistics,vehicle finegrained classification is of great significance to the intelligent transportation system.However,some vehicle types have high similarities in appearance and color.Their difference lies in the subtle features of the local region,which make it difficult for general classification algorithms to distinguish similar vehicles.Therefore,based on deep learning,this thesis proposes a finegrained vehicle classification algorithm in highway scenes by using the self-attention mechanism,feature relation reasoning,and subtle feature enhancement.This algorithm can correctly distinguish similar vehicles by mining subtle features between similar vehicles.In order to mine the distinguishing features between similar vehicles,based on the selfattention in vision transformer,this thesis designs a vehicle part selection module to locate and select the local subtle features of vehicles,which can reduce the interference of useless features in the classification process.In the meantime,for the highway scene,this thesis constructs a vehicle dataset,which covers a variety of road scenes.Through experiments,this thesis verifies that the vehicle part selection algorithm can improve the vehicle fine-grained classification task compared with the benchmark method.To utilize the feature relationship to infer the vehicle category,this thesis constructs a feature relationship matrix based on the similarity information and location information of the features.Next,a graph convolution network is used to fuse the relationship matrix and vehicle features to obtain the feature sequence with a spatial and semantic relationship.Simultaneously,according to the neighborhood relationship of features,this thesis proposes a vehicle feature enhancement algorithm.This algorithm enhances the features selected by the vehicle part selection module,thereby increasing the subtle information on vehicle features.Through experiments,it is proved that the classification accuracy can be further improved by establishing feature relationships and using the feature enhancement algorithm.Based on the vehicle fine-grained model proposed in this thesis,a vehicle classification system is implemented to classify the vehicle images containing a single target.In order to expand the system,the target detection algorithm is used to identify the vehicles in the video and count the number of vehicles.The system provides an operation page for uploading pictures and videos,and visualizes the results of model detection on the page.