Research On Vehicle Re-identification Method Based On Deep Learning Multi-model Fusion

Vehicle Re-identification(Re-ID)is a new technology researched in the field of smart urbanization and road traffic safety in recent years,which is used to exactly fit the characteristic of vehicles in numerous various surveillance cameras in an intricate traffic condition,and to judge whether the vehicle images shot by various cameras of a given target vehicle are the same target in different scenes.This effectively saves the workforce and financial resources required for accurate retrieval of massive vehicle data by traditional methods,and promotes the development of public security,unmanned driving,intelligent transportation and other related fields.Therefore,the task of rapid vehicle feature positioning and re-identification can be completed by the adoption of the computer vision technology to match the image features of target vehicle.In response to this issue,this paper integrates feature representation learning and metric learning to design a multi-model fusion vehicle feature extraction network,and use vehicle time-space information to reorder,so as to complete the task of vehicle re-identification with high accuracy.The specific research content of this paper is as follows:(1)Research on multi-branch network architecture to collaboratively extract vehicle global features.By analyzing and comparing various basic convolutional neural networks for feature extraction,a multi-branch deep learning network architecture is proposed to realize the representation learning of vehicle global features,and the improved triple loss function is used to optimize global feature recognition to enhance distinctive representation of the initial vehicle images.The experimental results illustrate that the multi-branch network learning is more targeted than single network learning in extracting global features of vehicles,and can effectively analyze vehicle global feature information to distinguish vehicles with larger appearance differences.(2)Research on vehicle local feature representation methods to extract part features.Propose an improved Faster R-CNN to extract detailed target features of vehicle parts,and increase the attention mechanism module to enhance the feature learning ability of the main representation.In this model,the cost-effective Res Net50 is selected as the basic feature extraction network;the FPN is selected to accurately detect and recall local vehicle parts;the Ro I Align is used to avoid the error problem caused by two consecutive quantizations,and can improve the pooling accuracy of the target suggestion frame.Finally,an attention mechanism module is introduced,which aims to enhance the target detection ability of the improved network.(3)Research the multi-model fusion network to complete the task of vehicle reidentification.By analyzing the extraction model of vehicle global features and local features,a multi-model fusion network based on vehicle feature mapping is designed.The network can use global features to distinguish vehicles with large differences,and based on fine-grained local features to strengthen the discrimination between similar vehicles.And the loss function is improved to optimize the generalization power of the entire network.At the same time,a new post-processing strategy for spatial-temporal re-ranking is introduced to reorder the initial search results.Experiments on public vehicle re-identification data sets verify that the model has obvious effectiveness and accuracy for the task of vehicle re-identification.