The development of a traffic sign recognition system

Abstract

This study focuses on the development of a traffic sign recognition system (TSRS) in which various methods applied in machine vision were employed to detect and recognise traffic signs in real-world images. The main aim of the TSRS developed was to develop a system which could be used by drivers, in normal weather conditions during daytime. The system focuses on the recognition of a limited number of 'warning', 'prohibition' and 'directional' signs, and proved to he successful in detecting and recognising signs which were predominantly red and blue, had different levels of illumination and those which had shadows cast on them. The TSRS developed was also successful in detecting signs which were slightly rotated or skewed, or which had their lower part occluded by other objects. The methods used focused on two main stages; the detection stage and the recognition stage. At the former stage, colour was the main determinant for detection. Colour segmentation was done by using HSI, Adaptive Histogram Equalisation and RGB in order to classify pixels into colours of interest. Run length encoding and a tree data structure were used in order to merge pixels together to extract possible regions of interest. 80% of the signs tested were successfully detected at this stage. At the recognition stage, the Wisard Weightless Neural Network was used. This method gave a success rate of 95% in sign recognition. Another method which focuses on shape edge analysis was also developed, with a success rate of 77%. The signs were finally classified into their target classes using the Wisard Classifier. The testing was done on two sets of images, with one set of images having better resolution than the other. The images with the higher resolution rate proved to be the most successful at the testing stage.