Traffic sign recognition in outdoor environments using reconfigurable neural networks

Abstract

A novel technique for recognizing street sign landmarks for mobile robot navigation is presented. Due to the motion of the mobile robot, the apparent target shape is distorted in terms of scale, occlusions, translations as well as rotations. The recognition is based on a self-organizing neural network called the reconfigurable neural network. This network also has the ability to online add new target patterns into memory thereby eliminating the need for retraining of the network. Update normalization is used during the training process to improve network stability. The learning rules can also be used to estimate the optimality of the training. The network has been successfully trained with street sign images which were subject to the various distortions.