Research On Traffic Sign Detection And Recognition

With the development of the road traffic network and the increasing number of vehicles,the demand for road traffic safety and intelligent transportation system is increasing day by day.As an important subsystem of intelligent transportation,traffic sign recognition system has been widely used in the fields of Driver Assistance Systems,intelligent navigation,autonomous vehicle,traffic sign maintenance,intelligent transportation and so on.For safety,traffic sign recognition systems must be efficient and accuracy,which makes highly demands to the traffic sign detection and recognition algorithm.In this paper,a series of solutions to the problem of traffic sign detection and recognition are proposed to meet the requirements of accuracy and real-time of traffic sign recognition system,as well as various difficulties and challenges in natural scene images.The specific contents include:(1)This thesis proposes a symbol-based traffic sign recognition method using two-layer image representation.In order to solve the multi-category classification problem of symbol-based traffic sign recognition with unbalanced inter-classes,we first extract and fuse features(e.g.,dense SIFT features,HOG features,and LBP features)and encode them with a k-means generated codebook and Locality-constraint Linear Coding.Then,each symbol-based traffic sign is represented by Spatial Pyramid Matching to maintain translation invariance and achieve multi-scale representation.Finally,based on this image representation with strong expressive ability and generalization ability,the symbol-based traffic sign classification problem can be easily solved by using a simple linear support vector machine classifier.The experimental results show that this method achieves the most advanced recognition rate of 99.67% and 98.77% on the German Traffic Sign Recognition Benchmark(GTSRB)and Belgium Traffic Sign Classification(BTSC)respectively at that time.(2)This thesis proposes a new end-to-end symbol-based traffic sign recognition framework based on deep learning.This framework consists of two parts: one is the symbol-based traffic sign proposal extraction part guided by the Full Convolution Network;the other is a deep convolution neural network classifier.Our core idea is to generate the symbol-based traffic sign proposal by using the object proposal method guided by the Full Convolution Network to achieve fast and accurate symbol-based traffic sign detection and recognition.This,on the one hand,reduces the search range of the object proposal method which enables it to produce a small number of distinguishing proposals quickly.On the other hand,the reduction of proposals also improves the accuracy and speed of the deep convolution neural network classifier to some extent.We have tested this framework on a publicly available symbol-based traffic sign benchmark: Swedish Traffic Signs Dataset(STSD).The experimental results show that,this framework obtained the most advanced results(accuracy rate 98.67,recall rate 93.2727)at that time.(3)This thesis proposes a new segmentation-detection framework for detecting text on text-based traffic sign,which consists of two kinds of deep learning methods.Text-based traffic sign has two characteristics,one is text on text-based traffic sign,the other is the size ratio of text and its corresponding text-based traffic sign is relative fixed.Making full use of these two characters of the text-based traffic sign,this method first applies a Fully Convolutional Network to segment the candidate text-based traffic sign areas,and then uses a fast neural network to detect text on the candidate text-based traffic sign areas.This,on one hand,reduces the search range of the text detector and removes the background text.On the other hand,it solves the problem that the text detector needs multi-scale input to a large extent,and improves the efficiency and precision of the text detector.A large number of experimental results show that this method not only obtains the most advanced results on the publicly available English text-based traffic sign dataset: Traffic Guide Panel dataset,but also shows good performance in our own collection of text-based traffic sign dataset: Text-based Traffic Sign Dataset in Chinese and English(TTSDCE)which contains both Chinese and English texts.This indicates that this method is a fast and effective method to detect text on text-based traffic signs for multiple languages.