Research Of Vehicle Information Recognition Technology Based On Nerual Network Learning

Recently,with the popularization of vehicles,people become more concerned about traffic safety,so the road traffic surveillance has arised.Moreover,with the development of information technologies,the traffic surveillance system has become automatic and intelligent.However,the circumstances in the real surveillance video are so complex that traditional methods fail to adapt it,which means a lack of robustness.In this paper,based on neural network theories,we propose the solution for vehicle detection,vehicle type classification and vehicle logo recognition,and verify the algrorithm feasibility with lots of experiments.The main contribution of this paper in vehicle information recognition are as follow:(1)Introduce the principle of vehicle detection,vehicle type classification and vehicle logo recognition;propose the architecture of vehicle recognition system.(2)Introduce the convolutional neural network theories,and propose a vehicle detection method based on Fast-RCNN.With the generalization of neural networks,the proposed method can adapt to serveral circumstances in the reality.We can get the feature maps by deep networks,and get the detection results by the fully connected networks.Additionally,in order to improve the detection rate in the night,we propose the detection algorithm in the dim light conditions.(3)Propose a novel vehicle-type classification method based on adaptive feature learning,to solve the weakness in the traditional methods.We obtain the main features by the convolutional neural networks,and obtain the weak labels by the extreme learning machine(ELM).Finally,we combine these two features proportionally,and obtain the vehicle-type feature cluster centers by K-means.Because of the adaptivity,this method can adapt to the various real circumstances(4)Propose the structure of vehicle-logo recognition system.First,we provide a ‘coarse to fine' method to locate vehicle logo.Then,the mathematical principle and model of ELM is proposed.Finally,a local receptive fields based ELM network is implemented to recognize the vehicle logo.In order to impove the roubustness,we improve the algorithms on the feature filters and network learning method.(5)Several experiments are performed to verify the feasibility and effectiveness of these algorithms,and we finally analyse the experimental results for the further research.