Research On Vehicle Type Recognition Method Oriented To Traffic Scene

To alleviate the great pressure in traffic management which brought by the increasing number of automobiles,Intelligent Transportation System(ITS)emerges as the times require.As a part of intelligent transportation system,vehicle type recognition plays an important role in many traffic scenarios,such as parking lot management,unmanned toll station,suspected vehicle identification,vehicle type query,unmanned driving and so on.With the increasing demand of traffic management,the high accuracy of vehicle type recognition is also increasingly demanded.According to the requirements of different traffic scenarios,vehicle recognition is mainly divided into coarse-grained classification and fine-grained classification.In coarse-grained classification,the main difficulty of vehicle type recognition is how to achieve high accuracy in the case of small sample,while in fine-grained classification,the difficulty of vehicle type recognition mainly lies in the intra-class difference and inter-class difference between vehicles.For this reason,this paper studies vehicle type recognition method oriented to Traffic Scene from two aspects: coarse-grained vehicle recognition and fine-grained vehicle recognition.The main work of this paper includes:(1)A method which improves the quality of complex scenes image is proposed.Firstly,the encoder-decoder fullly convolution neural network is used to acquire the foreground contour of the image,and then the local maximum method is used to obtain the single pixels of the contour.Finally,the accurate edge of the foreground vehicle is obtained,and the black pixels are used to cover the background to highlight the features of the foreground area.The validity of background removal is also verified in subsequent experiments.(2)Aiming to raise the low accuracy of coarse-grained vehicle type recognition based on small sample datasets,a two-stage vehicle type recognition method is proposed.The first stage is the background interference removal stage,in which the high-quality dataset extracted from the foreground is used as the training dataset.In the second stage,a multi-scale feature fusion method is proposed for data training.which integrates handcrafted features with learning-based feature.The combined features are input to Support Vector Machine which has strong generalization ability to train the recognition model.Finally,this method achieves 90.96% accuracy on the public dataset MVVTR better than other similar methods.(3)Aiming at the problem of small difference between classes and large difference within classes in fine-grained classification,a method of data augmentation combined with hybrid convolution model is proposed to improve the accuracy of fine-grained vehicle recognition.The dataset augmentation method is proposed to expand the image dataset while increasing the differences between classes and narrowing the differences within classes.Based on the expanded dataset,a hybrid convolution model is proposed to integrate the weak classification model of vehicle type recognition,which achieves the same effect as strong classification model and achieves 92.21% accuracy on the public dataset Cars196.