Research On Vehicle Brand Recognition Algorithm In Complex Scene

The vehicle recognition technique is an important part of intelligent transportation system.In particular,vehicle brand recognition can effectively assist vehicle retrieval and vehicle verification and is thus helpful to detect vehicle illegal acts and criminal vehicles.Therefore,the problem of vehicle brand pattern recognition in the traffic monitoring scenes has potential values for a variety of applications.In recent years,variety of algorithms has been proposed for vehicle brand recognition,but there exist two issues.One is that the used vehicle images are with fixed front angle captured by HD bayonet camera,and another is the limited number of vehicle types.In the real monitoring scenes,the images are usually recorded in more challenging systems,e.g.,bayonet imaging system and virtual bayonet imaging system.In such circumstance,the vehicle angles are complex and changeable,and there also has a serious imbalance problem of vehicle samples.We study the problem of vehicle brand recognition under complicated scenes in this thesis,and the major works are as follows.(1)We propose a novel method based on batch normalization for bayonet vehicles recognition.In the bayonet monitoring scenes,mass data are collected to train deep neural network(DNN)models,in which the training time is unacceptable big and the network is also difficult to converge.To handle this problem,we combine the data augmentation and batch normalization in DNN.Experimental results show that the embedded layer and the data augmentation greatly improve the recognition rate with much less training time and fast convergence speed.In addition,we also contribute a large-scale bayonet vehicle dataset with a hierarchical structure.(2)To handle the problem of non-equilibrium samples in virtual bayonet,a novel vehicle recognition method is proposed.In real scenarios in virtual bayonet,the vehicle samples are seriously unbalanced,using traditional DNN usually obtains bad performance.To this end,we add the focus loss in network training,and the classification accuracy of vehicle brands is effectively improved.In addition,by using a wider network module,the capability of feature selection and extraction are effectively improved,and we thus improve the recognition accuracy of vehicles under virtual bayonet scenarios.Furthermore,a virtual bayonet dataset containing 10192 vehicle images is constructed.(3)In practical complex scenes,the vehicle angles are changeable which seriously reduces the generalization and migration ability of vehicle recognition algorithms.To handle this problem,we propose a novel progressive CNN architecture.First,the vehicle angles are estimated via a backbone network,and then the vehicle brand of the corresponding angles are classified by branch networks.Moreover,the end-to-end network training and testing is designed and implemented.The entire vehicle identification framework can recognize the vehicle angle and brand in a unified framework and significantly improve the vehicle brand recognition accuracy comparing with baseline methods.