The Research Of Traffic Sign Recognition Algorithm Based On Tensor Flow

With the development of technology and the progress of mankind,intelligent driving has attracted more and more attention.One of the most important reasons is that machine has an excellent performance in reducing the influence of subjective factors,which is the cause of most traffic accidents.Beside this,intelligent driving is more comfortable and convenient.Traffic sign recognition system based on computer vison takes an important part in it.On the other hand,the theoretical research of deep learning algorithm has been developing rapidly in recent years.Related new technologies also have been applied to various fields.TensorFlow is an excellent deep learning framework that performs well in all aspects.This paper will use deep learning based on two-stage detectors which combine the advantages of traditional research methods and deep learning,exploring traffic sign preprocessing,detection and recognition algorithm.At last,we can solve this problem with high accuracy,high stability and high speed algorithm,and then build the system with TensorFlow.The main content of this paper are as follows: in image preprocessing,this paper adjusts the scale factor by adding feedback in Retinex algorithm.As a result,color distortion is largely reduced and the details are clearer.In terms of traffic sign detection,this paper proposes to extract candidate regions by combining HSV color space rough segmentation with adaptive detection algorithm.At first,this paper removes large amounts of redundant information quickly by HSV color space segmentation,and sets different priorities based on different colors.Then,this paper chooses different candidate area extraction algorithms according to the number of traffic signs in an image,which not only improves the accuracy,but also reduces the consumption time and optimizes the performance.In terms of traffic sign recognition,this paper draws on and improves the structure of AlexNet deep convolution neural network,solves the problem of image normalization,reduces network parameters,optimizes the loss function,and ultimately improves the accuracy and speed of recognition.All of the algorithms above are tested on the self-built traffic sign library or German Traffic Sign Recognition Benchmark(GTSRB,German traffic sign library).Finally,the whole system is implemented based on TensorFlow,and the algorithm is verified by the collected video in the traffic recorder.The results show that the algorithm is practical,and has excellent performance in accuracy,stability and real-time.