Using the data samples described we attempted several learning methods to detect traffic lights in grayscale image, such as: Multi Layer Perceptron, genetic, AdaBoost, etc. We also tested different features like: control points, connected control points [12], or haar features [13]. In this paper we will only detail the cascade classifier training with Adaboost and Haar features since this fast method gives good results for most of the objet-recognition tasks [13]. A future paper should be written to detail the exhaustive performances of each learning process.To apply cascade classifier to the traffic light recognition problem we used the sets detailed in Table I. Before training, brightness of each sample was normalized. Each cascade classifier contains at least 9 boosting classifiers which lead to a big computation time. The cascade classifiers trained are described below in Table II and the evaluation of the training is detailed in Section V.