Sistema inteligente de visión artificial aplicado a la detección de señales de tráfico para su aplicación en vehículos

Resumo

This thesis addresses the problem of the design and development of a traffic sign recognition system. The signs to be detected are the circular ones, either prohibition and obligation ones, the triangular hazard-warning sign, and the give way sign. The work is based on a monocular vision system with a monochrome camera onboard the vehicle, capable of operating both day and night and in different weather conditions. A sign-recognition full system is divided in three stages: detection, tracking and classification. In this work, a detection stage based on shape-analysis is proposed. A novel technique for circular and triangular sign selection is applied from the codification of the contours obtained by applying Canny-operator to every contour, with adaptive threshold level depending on the gradient magnitude. This technique enables overcoming, to a great extent, undesired partial occlusion or contour merging. Finally, detection is carried out by applying the Hough transform (HT) to selected contours; HT for circles is used if working with a circular sign, while HT for straight lines is applied to triangular signs. The HT is performed estimating the search-parameters of the figures, so that processing time is considerably reduced. In order to implement the classification stage the use of two support vector machines (SVM) have been proposed, for circular and triangular classification respectively. The training data base has been built by performing different transformations on some artificialsample signs, such as rotations, translations, scale changes, etc. Besides, different kind of noises were added so that any possible effect in a traffic actual situation is considered in simulations. Using ROC curves, commonly used for viewing and comparison of results, it has been determined as the best model of training as the kernel and the optimal cost function. Once the best classification strategy is chosen, the tracking stage has been designed making use recursive filters, similar to kalman filter, associated to each sign-candidate; this makes spatial validation easier. Finally, a time-validation module is proposed, based on a probabilistic approach. Finally, the system has been tested in real traffic conditions, with different lighting conditions, with good results. Furthermore, this work has been used as a commercial system based on inspection of traffic signs for thousands of miles with excellent results.