Study On Traffic Sign Detection And Recognition

ABSTRACT:Due to advances in technology, road safety has been improved tremendously and efforts are now focused on problems caused by the driver’s error. It has been well-known that many traffic accidents occur due to the drivers miss or ignore traffic signs. With the attention of drivers getting diverted due to distractions like cell phone conversations, missing traffic signs has become more prevalent. Also, poor weather and some unfriendly driving conditions which make the drivers not to alert all the time and see every traffic sign clearly on the road. Besides, most cars do not have any form of traffic assistance. Traffic sign detection and recognition system can provide information for the drivers in time, which allows the driver to make timely responses to unexpected situations, thereby avoiding or reducing traffic accidents. This thesis investigates traffic sign detection and recognition technology, including image preprocessing, detection, feature extraction and recognition. The main works of the thesis are as follows.(1) In the stage of image segmentation, the MSER algorithm is used and combined with the judgment of geometry features, such as sizes and scales of traffic sign. This method can effectively segment the traffic sign candidates while excluding a large number of non-traffic sign regions.(2) For the purpose of detecting of the traffic signs, an algorithm based on multi-features and high-credibility region is proposed. Firstly, the color, shape, and texture features are extracted and classified with SVMs to find the traffic signs. Then, an adaptive iterative adjustment based on credibility function is used to obtain the optimal traffic sign area from the detected overlap candidates, which may be in the presence of offset and not matching with real traffic sign region. The result shows that the recall rate of the proposed algorithm on GTSDB and STSDB achieves97.1%and98.5%, and the average coverage is over94%, which exceeds most traffic sign detection algorithm at present.(3) A method of traffic sign recognition based on local features and cascade classifiers is proposed. Firstly, traffic signs are classified into four categories, including circle, triangle, diamond and upside-down triangle, based on local shape features and expert committee. Secondly, the internal components of circular and triangular traffic sign are extracted and recognized based on a method of similar-template matching. An advantage of the method is that it does not require a lot of features and samples. The algorithm is tested on2601traffic sign images in GTSRB database. The accuracy of recognition reaches98.3%and is higher than deep learning, HOG+SVM and HOG+Random forest, which extract the global features of traffic signs in training and testing.