Sistema de detección y reconocimiento de señalización en carretera mediante técnicas de procesado digital de imagen e inteligencia artificial

Abstract

This Ph.D. thesis nds itself in the context of Intelligent Transportation Systems and its main objective is to carry out a study about dierent techniques in the eld of digital image processing and computer vision applied to automatic traffic sign recognition. Even when multiple number research groups dedicated to traffic sign detection had arisen in the last decade, most of them have been concerned with an only stage and their experiments have been developed with reduced sets. The rst research line of the thesis focuses on the study and development of a complete traffic sign detection and recognition system, which can be used for road maintenances and for intelligent vehicles. The system is based on a modular structure with four stages that correspond to segmentation, shape classication, idegram recognition and tracking. Although the dataset we built in this thesis contains all Spanish categories, the system is easily adaptable to the signs of another countries. The good properties of generalization exhibited by SVMs lead us to introduce them as a statistic alternative, specically in the shape classication and ideogram recognition stages. However, the process of recognition with SVMs is slow since the system must manage many classes and requires a lot of training data with different image conditions. It demands a high computational complexity and supposes a bottleneck in the complete system. The ideogram recognition task is structured in three steps: preprocessing, descriptor extraction and classication. The second research line in this thesis focuses on the optimization of the ideogram recognition by making a comparative study between existing parametric algorithms. Thus, we search the best architectures and descriptors from the points of view of computational cost and overall performance. With the aim of reducing the complexity at the recognition stage, we have proposed a novel technique based on the search of spatial distribution of ideograms by finding common similarities that can be shared across the classes. The procedure employed in this work allows us to cluster categories with similar spatial distribution. In this way, the number of potential classes is reduced in the sense that classification process only takes into account the classes grouped within the selected cluster. In addition, an adaptive descriptor is extracted independently in each cluster, where the number of features of each region depends on the information measurement. Thus, it is possible to control the computational cost at the recognition stage while keeping the overall performance of the system. Our experimental results show substantial time reduction with respect to the conventional procedure when all classes are considered in the multi-classication problem. The proposals described in this Ph.D. thesis have been assessed for a Spanish traffic sign dataset. This has facilitated the development of the experimental stage, while achieving signicant results that support the viability of the use of the methods presented. As there does not exist a public standard dataset, we have created one with all the Spanish traffic signs. In order to include scenes and samples under critical situations, the sequences have been acquired by varying the configuration of cameras in different scenarios.