Research On Vehicle Attribute Recognition Method In Traffic Monitoring

As an important member of the intelligent transportation system,the effective identification of various attribute information of vehicles in traffic monitoring can improve the operation efficiency of the intelligent transportation system.The traditional vehicle attribute recognition includes only a single attribute such as the type or color of the vehicle,which can no longer satisfy the existing traffic system.In order to improve the reliability of vehicle detection and positioning in actual monitoring,a deep neural network is used to establish a model that can identify various vehicle attributes in traffic monitoring,including four attribute categories:vehicle type,vehicle color,vehicle logo and license plate.The main research contents of this article are as follows:1.Analyze and research the research direction and progress of vehicle attribute recognition and deep learning at home and abroad,and summarize the shortcomings of various vehicle attribute recognition methods.2.In view of the problem that the current vehicle attribute recognition has a single attribute and cannot meet the current traffic needs,an improved network based on the YOLOv3 network is proposed.For the size of each attribute of the vehicle,a hierarchical training method is adopted to reduce the depth of the YOLOv3 network.Retain part of the convolutional layer,and then add a pooling layer after each convolutional layer to extract feature of the vehicle attributes.Finally,the output multiple scales are stitched together to achieve the fusion identification of the various attributes of the vehicle.3.Establish a vehicle multi-attribute data set that meets the traffic monitoring environment,and use this data set to prepare the network for model training to complete the training and test verification of the vehicle attribute recognition model.The experimental results show that:(1)The vehicle attribute recognition model based on the improved YOLOv3 network used in this paper can realize the fusion recognition of multiple vehicle attributes,and overcome the single attribute problem of the existing vehicle recognition methods;(2)The use of hierarchical training of various vehicle attributes effectively avoids the problem of extended recognition time due to the increase of network depth,and effectively improves the accuracy rate under the condition of meeting real-time performance;(3)Test and verification under various monitoring scenarios,all show good applicability.It is suitable for the attribute recognition of vehicles in traffic monitoring.