Research On Fine-grained Vehicle Identification In Complex Contexts

Vehicle type recognition technology is an important component of intelligent transport systems.Currently,many research results of vehicle type recognition technology have been applied to real traffic scenarios,which have reduced vehicle management costs and improved the efficiency of traffic enforcement.However,most vehicle type recognition technologies can only perform coarse-grained recognition of a small number of categories such as large trucks,cars and SUVs,and these large categories of data do not provide more effective information.In addition,complex backgrounds such as bad weather,dark scenes,trees,and different vehicle occlusions can increase the difficulty of vehicle feature extraction,which in turn affects the accuracy and adaptability of vehicle recognition algorithms.To address the above issues,this thesis explores fine-grained level vehicle recognition techniques in complex backgrounds based on deep learning network models,and the main work is summarized as follows:1.A new fine-grained model for model recognition is constructed based on the Res Net network,the attention mechanism,and the encoder module.The Res Net network is modelled with CNN with good inductive bias,but its ability to model long sequences is weak compared to the attention mechanism.To enhance the feature extraction capability of the residual network,channel attention and spatial attention are added to the input and feature output positions of the residual network.In addition,to further enhance the modelling capability of the residual network,two encoders based on the attention mechanism are added after the final pooling layer.Three sets of comparison experiments demonstrate that the attention mechanism module and the encoder module effectively improve the classification accuracy of the model.In particular,the classification accuracy of the Res Net 50 network with the addition of the above two modules reached 85.69%,which was 2.43 percentage points higher than that of the original Res Net 50 network.2.A YOLOv4 fine-grained vehicle detection and recognition model with a CBAM-Res Net50 front-end network is proposed based on the Res Net 50 network and the attention module to improve the YOLOv4 model.The CBAM-Res Net 50 network is built by first adding the spatial attention and channel attention modules to each residual block of Res Net 50,and then replacing the backbone network CSPDar KNet53 in the YOLOv4 model with the CBAM-Res Net 50 network built in this thesis.The final experiments show that the YOLOv4 model with the CBAM-Res Net 50 front-end network is faster,requires fewer parameters and achieves a m AP value of 99.02%,an improvement of 1.16 percentage points over the original YOLOv4 model.3.The Stanford 196 and Comp Cars datasets used for the experiments were pre-processed separately.The number of training sets in the Stanford 196 dataset was expanded using data augmentation,and all model images were cropped to a size of 224 × 224 to improve the generalization ability and robustness of the network.A Comp Cars-based multi-category car dataset was constructed by collating 116 categories of car models from the Comp Cars car model dataset,each with more than 150 images,and cropping all the images to a size of 448 × 448.