Localização e reconhecimento de placas de sinalização utilizando um mecanismo de atenção visual e redes neurais artificiais.

Abstract

The main objective of this dissertation is to investigate the problem of the traffic signs detection and recognition, using data from a video camera attached to a moving car. We designed an architecture containing two main modules: a detection module to automatically locate traffic signs inside each frame; and a recognition module to classify the located regions based on a set of previously trained images. For the detection module we are using a saliency-based attention mechanism (bottom-up), which is constructed from a Gaussian Pyramid, and locally oriented neighborhood operators. Some initial tests of this mechanism showed promising results since signs were present in most image salient regions. Preliminary experiments with the recognition module presented good results, with 84.40% recognition average rate. The results also indicate that better rates could be reached if we increased the number of examples in the training set. However, when using the outputs of the Detecion Module, in which sign images are not necessarily centred, the use of a monolithic neural clasifier presented, for all classes, insatisfactory results. In face o f this problem, we developed new simple experiments involving binary classification networks (discriminating between 2 classes at a time). These experiments have shown that it would be possible to employ a classification strategy combining the output of these binary networks. Results have also shown that better recognition rates could be achieve though an increase in the training set size.