Research On Traffic Video Detection And Vehicle Classification Based On Deep Learning

With the increase of car ownership,the traffic problems become more and more serious.At the same time,deep learning has been rapidly developing,and has brought great changes to the pattern recognition;especially it also provides some new solutions to kinds of research areas.Therefore,the application of deep learning to solve traffic problems has become a research area.The convolution neural network method of deep learning is used to implement traffic target detection and vehicle classification in traffic video in this paper,and to provide technical support for intelligent transportation system in the purpose of alleviate traffic congestion and so on.The main contents of this paper are as follows:Firstly,the theoretical model of deep learning is introduced,which is mainly divided into deep belief network,stacked self-encoder and convolution neural network.Three things are mainly studied here,including the composition of convolution neural network,the features of convolution neural network different from the traditional neural network,and the training mechanism of convolution neural network.In this paper,the convolution neural network is used to extract the feature automatically,because the process of expressing feature design is complex and the scope of adaptation is limited.Based on the region-based convolution neural network(RCNN),a traffic video detection scheme is designed,which combines the advantages of the Fast RCNN framework and the RPN network.According to the characteristics of the traffic target contours,this paper improves the shared convolution network in the traffic video detection network,and mainly deepens the convolution network from 5 layers to 13 layers,resulting in the increase of mean Average Precision(mAP)of traffic targets by around 3 percentage points.In the existing vehicle classification method,the vehicle is just roughly classified,which is unable to meet the need of the Vehicles Internet.The deep residual network model for vehicle classification is used in this paper,with 64 vehicle brands.In the process of designing the model of classification network,the performance of the commonly used image classification convolution neural network is compared,and the deep residual network is chosen as the main frame of the vehicle classification network.The accuracy of vehicle classification network can be improved by 97.3% on the CompCars dataset and by 89.4% on the VehicleID dataset,which verifies the validity of the vehicle classification network.Finally,the traffic detection network and the vehicle classification network are tested in the image and video of traffic videos.The detection network can obtain high detection rate in different conditions,such as sunny day,night,rainy day and traffic congestion.The detection rate can reach 98.7%,and shows good robustness.The classification network achieves a top-5 accuracy rate of up to 88% in a vehicle test set based on videos.The experimental results show that the detection network and classification network designed in this paper have certain practical values.