Vehicle Re-identification Based On Multi-attribute

The development of artificial intelligence is constantly improving the level of intelligence in the transportation field.As part of the intelligent transportation system,vehicle re-identification refers to:giving a vehicle image taken under a fixed camera as an inquiry object,and vehicles images under other cameras.A gallery is composed of images,and an image of the same car as the query object is found in the gallery.The traditional machine learning method performs vehicle re-recognition by extracting manual features and classifiers.The amount of data required for training the model is small and the hardware requirements are low.When the vehicle image is sufficient and the hardware conditions are suitable,the depth-based nerve is used.The feature extraction ability of network extraction is stronger than the manual feature.The vehicle image of the vehicle re-identification dataset is often affected by the camera’s viewpoint,illumination,occlusion,and resolution.This paper explores the vehicle re-identification algorithm based on the real road vehicle dataset,and demonstrates the effect of the vehicle re-recognition algorithm through the web platform.The main work and innovations of the article are as follows:(1)In order to solve the problem of images’s excessive difference between different angles of the same car,this paper proposes a vehicle re-identification algorithm based on vehicle angles.We divide the vehicle dataset VeRi-776 according to vehicle color and models,then treat each image as a class,extract vehicle features through ResNetl8 model,and t-SNE dimensionality reduction,K-means clustering,"elbow method" selects the value of the number of centroids K to get detailed information of angles.Then,using the information of angles as an aid,the vehicle feature extraction based on the vehicle ID tag classification is performed by the ResNet50 model,and the discriminative vehicle feature is obtained.Finally,the effectiveness of the algorithm is proved on the VeRi-776 test dataset.(2)In order to solve the problems caused by resolution and occlusion,this paper proposes a vehicle re-identification based on vehicle color and models.We design two modules.In the vehicle feature extraction module,we modify the deep neural network to make it have a multi-branch structure,and use global branches,BN branches,and regional branches to extract and classify the vehicle features.In the attribute recognition module,we use the vehicle color and models attributes to change the vehicle attributes and the vehicle attributes as the auxiliary vehicle attribute classification.The attribute information is used to assist the vehicle feature extraction during training,and the obtained depth features have strong robustness and discriminative power.Finally,the effectiveness of the algorithm is verified on the VeRi-776 test dataset.(3)This paper constructs a vehicle re-identification demonstration system based on web platform.This system takes the vehicle image incoming by the user as the query object,and then calls the vehicle re-recognition algorithm to find an image similar to the query object in the vehicle gallery.We use MySQL to manage system data,design controllers and web view templates through the ThinkPHP framework.The system architecture is divided into three modules:the logic layer uses ThinkPHP’s background controller to define the operation method;the data access layer uses MySQL to build the system database;and the user interface layer uses Bootstrap to build the user interface quickly and efficiently.