Invariant Road Sign Recognition with Fuzzy ARTMAP and Zernike Moments

In this paper, a novel approach to recognize road signs is developed. Images of road signs are collected by a digital camera mounted in a vehicle. They are segmented using colour information and all objects which represent signs are extracted, normalized to 36times36 pixels, and used to train a fuzzy ARTMAP neural network by calculating Zernike moments for these objects as features. Sign borders and pictograms are investigated in this study. Zernike moments of sign borders and speed-limit signs of 210 and 150 images are calculated as features. A fuzzy ARTMAP is trained directly with features, or by using PCA for dimension reduction, or by using LDA algorithm as dimension reduction and data separation algorithm. Two fuzzy ARTMAP Neural Networks are trained. The first NN determines the class of the sign from the shape of its border and the second one determines the sign itself from its pictogram. Training and testing of both NNs is done offline by using still images. In the online mode, the system loads the fuzzy ARTMAP neural networks, and performs recognition process. An accuracy of about 99% is achieved in sign border recognition and 96% for Speed-Limit recognition.