Traffic Sign Detection And Classification Method Research Based On Machine Learning

With the acceleration of urbanization in China,the number of car in cities has been increasing year by year,and travel safety has been the focus of the public's consideration.In order to lighten the driver's attention burden,machines with Advanced Driving Assistance System have been invented,and traffic sign recognition is one of the key functions of these devices.The traffic sign detection and classification method based on machine learning has the advantage of low cost,high accuracy and timely change of road conditions,which is the focus area of machine vision research.At present,the traffic sign recognition algorithm in real scene still has the problems of high miss rate and low recognition accuracy.The research work of this thesis is carried out around these problems.The details are as follows:1.Aiming at solving the problem of lacking proper traffic sign dataset,this thesis collects images from network and extracts frames from recorder video,then makes CTS64 dataset and its subset based on real scene in China.These dataset are used to train and evaluate algorithms,which help expand training dataset to some extent.2.Aiming at solving the problem that the common feature descriptors have high miss rate in the process of speed limit sign detection,a multi-scale pooling aggregated channel feature based on feature combination is proposed.The proper combination of channel feature at three scales,with excellent classification ability of AdaBoost cascade classifier,sharply improves the extraction skill of descriptor and effectively reduces miss rate.3.This thesis proposes a method based on digital image processing for feature extraction of speed limit sign.It is simple and effective in extraction.A variable threshold calculated from quadratic curve fitting replaces a fixed one,brings about high accuracy of extracting red pixels.The extraction algorithm based on region growth effectively segments the digital from background,and improves the robustness of feature extraction under the condition of sign tilting.Finally slice projection reduces the dimension of feature vector and reduces the complexity of multi-layer perceptron classifier.4.This thesis proposes a method based on multi-stage feature combination for traffic sign recognition.For the reason of numerous category and intricate shape of signs,an improvement is applied to VGG16 model on the basis of convolutional neural network.Convolutional layer's low-level features are combined with high-level ones along dimension axis,to improve descriptive power of model.The experimental results shows that 2.38% improvement was made in detection precision,2.17% in recall and 1.75% in classification accuracy from the CTS64 dataset.While from the famous GTSDB dataset,4.54% and 5.16% improvement were made in detection precision and recall respectively.