A Research Of Traffic Sign Recognition Based On Deep Learning

Road traffic sign is an important kind of road facilities,which contains plenty of traffic information,it plays an significant role in guiding the driver to drive safely.Unfortunately,the driver usually ignore the traffic sign on the road owing to their carelessness or fatigue driving,which cause disastrous tragedies and painful damage to the family and the society.Traffic sign recognition system is a critical component of advanced driving assistance system.It helps the driver to have a good grasp of the surrounding road situation in advance and reminds them to make correct response according to different situation,which greatly alleviates the driver's driving pressure and contributes to a sharp drop of traffic accident.However,the traffic sign recognition research meets tremendous challenges due to the complexity of real road,such as the changeable light intensity,terrible weather,shelter and so on.This paper utilizes the relevant machine learning and deep learning knowledge focusing on the traffic sign recognition algorithm improvement on the basis of summarizing a large number of domestic and foreign related research literature.The work of this paper is divided into two stages: traffic sign detection,which refers to detecting the candidate regions that contains traffic sign from the input image,and traffic sign classification,which refers to the classification of the detected traffic sign regions,the result of detection phase,and outputs the specific category label.The main work and innovation of this paper are as follows:(1)For the extraction of traffic sign candidate regions in the detection stage,this paper combined HSV color segmentation with Maximally Stable Extremal Regions algorithm to complete the rough extraction of traffic sign candidate regions.We extracted a large number of different color pixels from the training data set to obtain the ranges of the three color components(H,S,V)in HSV traffic sign images and determine the threshold of color segmentation.Because of the uniformity in hue and saturation and the more brightness over surrounding background of the traffic sign pattern,the traffic sign is in the form of connected domain after color segmentation.We utilized MSER method to implement candidate region extraction after the completion of color segmentation.Compared with the traditional sliding window method,our candidate extraction method improved a lot in speed.(2)For the feature extraction of the candidate region in the detection phase,this paper conducted a focused research on the relationship between the numbers of bins in different intervals of the histogram of oriented gradient feature and proposed an improved HOG feature according to the traffic sign's characteristics.Thus,the HOG feature extraction was simplified and the dimension of the HOG feature vector was effectively reduced,which contributed to an tremendous reduction in the detection time while maintaining the high-level precision.(3)In the classification stage,we inserted a color transformation module and a space transformation module into the conventional CNN and replaced the fully connected layers with extreme learning machine classifier to complete the classification output.Color transformation module was used to convert the original image into an appropriate color space.Different from the traditional color space transformation,it is a dynamic learning way to update the color transformation matrix.The spatial transformation module was used to apply space transformation operation on the feature map or original image,which helped to obtain the high-relevant regions for classification task.Extreme learning machine classifier has many advantages such as less parameter,faster,good generalization ability,which just makes up for the shortcomings of the fully connected layer of traditional CNN.(4)We made a small-scale Chinese traffic sign dataset,which contains 600 pictures and 963 traffic signs,to further verify the proposed detection method has good generalization ability.Every picture had been labeled.The proposed detection method in this paper were tested in the German traffic sign detection and further verified in self-made dataset.The proposed classification method was tested in German traffic sign recognition benchmark dataset.Experiments proved the proposed methods achieved a competitive result in accuracy and speed.