Research On Truck Brands Recognition Based On Images

With the development of expressway in China,the number of motor vehicles and the traffic flow of expressway are increasing.The traditional manual law enforcement and manual toll collection methods in expressway scenes are far from meeting the requirements of intelligent high-speed development.With the development of pattern recognition methods and GPU technology,vehicle detection and recognition based on computer vision has developed rapidly,which provides a solution for intelligent management of vehicle information and non-parking electronic toll collection in expressway.At present,the electronic toll collection system(ETC)for cars is becoming more and more popular,which improves the traffic efficiency of expressways.However,the intelligent toll collection system for trucks is still in the exploratory stage.In view of this,this paper studies the image-based methods of truck brand recognition,so as to provide technical support for intelligent toll collection system of trucks in highway scenes.Firstly,the characteristics of expressway surveillance truck image are analyzed,and the detection method of truck face based on Deformable Part Model is studied.The method is also compared with 4 other detection methods,which are Histogram of Oriented Gradients and Support Vector Machine,Local Binary Pattern and Support Vector Machine,Maximally Stable Extremal Regions(MSER)based license plate detection and vehicle symmetry,Edge Detection based license plate detection and vehicle symmetry.The results show that the performance of the method based on Deformable Part Model is better than the other four methods,and the detection rate reaches 98.02%.Secondly,five feature extraction methods are analyzed,which are Histogram of Oriented Gradients,Local Binary Pattern,Scale Invariant Feature Transform,Pyramid Histogram of Oriented Gradients and Edge Oriented Histogram.Based on Deformable Part Model and the above five features,comparative experiments of truck face detection are carried out.The experimental results show that the detection performance of Deformable Part Model based on Edge Oriented Histogram is better than the other four methods.The detection time is increased to 431 ms when the detection rate reaches 97.86%.Based on the surveillance truck images on the expressway,the image set of the truck face of Southeast University is constructed.Then,the feature extraction methods of Histogram of Oriented Gradients,Scale Invariant Feature Transform and K-Means clustering algorithm are compared and analyzed.Five kinds of kernels used in Support Vector Machine are analyzed,including Linear Kernel,Gaussian Kernel and Sigmoid Kernel,Polynomial Kernel and Laplace Kernel.It is shown that the performance of the method based on HOG feature and Linear Kernel function Support Vector Machine is better than the other four kernel functions,and the recognition rate reaches 89.63%.Finally,the AlexNet network model,Inception V3 network model,Xception network model and DenseNet-201 network model are compared and analyzed.The truck brand classification methods based on AlexNet-MMD network model,Inception V3-MMD network model,Xception-MMD network model and DenseNet-201-MMD network model are constructed.The experimental results show that the accuracy of Inception V3-MMD model is better than that of the other three models,and the recognition rate of truck brand reaches 94.85%.Based on Inception V3-MMD,Xception-MMD and DenseNet-201-MMD,a truck brand classification method based on transfer learning fused deep network model is proposed.The experimental results show that the truck brand classification method based on the proposed model is effective.The performance is better than Inception V3-MMD,Xception-MMD and DenseNet-201-MMD.The recognition rate reachs 99.07%.