Intelligent Traffic Elements Detection And Recognition Based On Deep Learning

Intelligent traffic is the indispensable part of intelligent cities and the most important development direction of the furture transit systems.In this paper,the detection and recognition of the three most important elements of intelligent traffic in real scene:license plate,vehicle type and traffic sign are discussed.This paper focus on the problems of intelligent traffic elements detection and recognition.It is difficult to train unbalanced data,the low robustness in the process of detection and recognition,model selection is difficult,small object detection and recognition effect is poor.Then a number of solutions are proposed in the aspects of visual features,model theory,and object detection framework and so on.The main works of this paper include:(1)Aiming at the problem that the rate of license plate detection is slow in HRV and the drawbacks of low robustness of license plate detection under the complex situation in monitoring scene,this paper proposes a type of motion detection method based on Epanechikov kernel function and frame-adaptation.On this basis,proposes a slide-frame algorithm based on AFRD,which can improve the processing speed on license plate detection without reducing the accuracy.In allusion to the drawbacks of low robustness of license plate detection under the complex situation in monitoring scene,proposes a license plate detection algorithm based on morphological,color and MSER,which makes the license plate detection more robust.Aiming at the problem that the uniqueness of the feature item and the low accuracy in license plate character recognition,we proposes an model combine the explicit feature classifier with implicit feature classifier,on the basis of this model,a new model based on the feature integration of explicit feature classifier and implicit feature classifier is proposed,which improve the accuracy of license plate recognition to some extent.(2)In view the problem of the imbalance of data in deep learning,the number of classes may vary greatly,and then affect the effect of model and the difficulty of model selection,in this paper,a Region-based Fully Convolutional Networks based on data equalization strategy(BR-FCN)is proposed on the basis of R-FCN,through a balanced data set to improve the effect of the model,and put forward the robust equilibrium strategies of Region-based Fully Convolutional Networks(BRR-FCN),further improve the effect of model through the integration of different frameworks and models.(3)For most of the target detection framework is not good for the detection of small objects,This paper is based on the idea of space in pyramid,and pyramid’s image based on the Faster-RCNN is proposed for traffic sign detection and recognition,and the feature pyramid is applied to traffic sign detection and recognition,which further improves the detection ability of the model to small objects,and improves the effect of the model.Experimental results on multiple benchmark datasets and real scene datasets show that:The method proposed in this paper can effectively improve the robustness of object detection and recognition,improves the effect of object detection and recognition,and reduces the difficulty of model selection.To sum up,the methods proposed in this paper effectively improve the indexes of object detection and recognition.