Applying a visual attention mechanism to the problem of traffic sign recognition

Driving a vehicle is a highly intensive visual information processing task in which traffic sign recognition plays an important role. Reports have shown that a great deal of the crashes at intersections and head-on collisions could be avoided if the driver had an additional half-second to react, and that inattentive drivers are the cause of most crashes. Therefore, this is an interesting field for the investigation of computer vision techniques. Within this context we are concerned with the automatic detection and classification of traffic signs in images acquired from a moving car. In order to reduce the amount of information to process, we employed a bottom-up visual attention mechanism to locate only the most promising points within each frame. Given a set of interest points, another module tries to match previously learnt traffic sign models against image regions centred on these points via a neural network approach. This paper focuses on the design aspects and preliminary results of the attention mechanism.