Research On Large-Scale Traffic Bayonet Vehicle Retrieval Method Based On Deep Learning

In recent years,with the rapid growth of car ownership in China,the road traffic safety situation is facing increasingly severe challenges.The traffic management department obtains vehicle images by installing more and more bayonet cameras and other equipment at highways and city intersections,and then combines with the intelligent traffic system analysis in the backend to increase the monitoring and management of vehicles.Vehicle retrieval,also known as "Searching Car by Car",is one of the most important components of an Intelligent Transportation System(ITS).Therefore,the research of vehicle retrieval technology has important application value.Due to the rapid increase in the number of vehicle pictures and the increasing number of vehicle categories,the traditional retrieval method has been unable to meet the retrieval requirements of large-scale vehicle images.With the rapid development of deep learning technology,convolutional neural networks have achieved excellent performance in the field of image processing.And the deep learning model can complete feature extraction quickly,with higher flexibility and universality.Therefore,this paper aims at the large-scale bayonet vehicle image dataset,using the deep learning framework to improve on the classic deep network model,and proposes a vehicle retrieval method that can be applied in practical scenarios.The contributions and main work of this paper are as follows:(1)The bayonet vehicle datasets CarsDataset-1 and CarsDataset-2 are constructed.Since the current vehicle dataset has fewer categories,smaller scales,and simple labeling information,we built a bayonet vehicle dataset with a scale of 150,000 images.In addition,we have labeled the data set with attribute information such as brand,type,style,age,and color.(2)A vehicle retrieval method based on multi attribute depth feature fusion is proposed.Traditional content-based vehicle retrieval methods often use low-level features(such as colors,shapes,textures,etc.)to represent images.However,this method is difficult to overcome the interference of background,illumination,and angle in complex scenes.The vehicle license plate-based vehicle retrieval method cannot effectively solve the problems of the fake plate and the license plate occlusion.At the same time,the feature based on the deep classification network have achieved good results in the retrieval problem.However,the existing method based on deep learning only constructs image pairs or triplet images based on the similarities and differences of the labels,and does not fully explore the information of the vehicle label and the model is difficult to train.In the label information of the vehicle picture,the vehicle type and color attributes are relatively independent and the number of categories is small.Therefore,based on the classification feature of convolutional neural networks,this paper proposes a method of fusing vehicle attribute(color attribute)features and classification features,and uses the fused features as retrieval features for vehicle retrieval.Finally,experiments on the large-scale bayonet vehicle dataset demonstrate the effectiveness of vehicle retrieval algorithms based on multi attribute deep feature fusion.(3)A vehicle retrieval method based on multi layer feature fusion and deep hashing is proposed.The features of the existing vehicle retrieval methods are generally floating-point features and cannot cope with the rapid retrieval of large-scale datasets.Therefore,this paper uses a hashing method for large-scale image retrieval.Compared with the traditional hash method,Hash method based deep learning can learn image representation and hash coding at the same time,and can achieve better results.Based on the end-to-end deep hash algorithm framework,this paper fuses the multi layer features of the convolutional neural network and then performs hash coding learning.Finally,Extensive experiment on the large-scale bayonet vehicle dataset demonstrates the effectiveness of the proposed method.