Abstract
In this paper a computer system for recognition of the circular prohibition and obligation road signs is proposed. Its main purpose is to assist a driver with information on passing signs, which in connection with an active cruise control can prevent dangerous traffic situations. Thus, the system can help to increase safety on our roads. The proposed system consists of two main parts: a detector and a classification module. Both employ soft classifiers. The detector does colour segmentation with a support vector machine, operating in a one-class mode. Then the circular shapes are found and passed to the classifiers. The classification module is built in a form of two committee machines, each composed of a series of expert neural networks and an arbitration unit. The two machines has the same internal structure, however they operate in different input spaces. The first one works in the spatial domain, which allows very accurate assessments of the relative vertical and horizontal shifts. The second machine operates in the log-polar representation which has the ability to represent rotations as vertical shifts. Each expert of a committee machine is realized as a Hamming neural network trained with affinely deformed set of reference road signs from the data base. Selection of a single answer from a group of experts is done by an arbitration unit which operates in the winner-takes-all mode. Additionally, arbitration has been endowed with a group support mechanism which boosts answers from a group of unanimous experts. The proposed system shows very accurate and fast response on circular road signs encountered in real traffic scenes. This has been verified by experiments which results are also presented and discussed in this paper.