Aplicación de técnicas de detección de objetos en imágenes mediante Deep Learning para la ayuda en la conducción en situaciones de tráfico complejas

Abstract

Machine learning is a discipline of artificial intelligence that has gained importance in recent years, becoming a key clement in a multitude of research projects. It has been, among others, the great enabler of autonomous driving. Autonomous vehicles, which are still under development, are those that can drive themselves without human intervention. For this purpose, they rely on the so-called ADAS, advanced driving assistance systems, with which they perceive the environment to make decisions. The hypothesis that has guided this doctoral thesis has been to use deep learning techniques to develop new assistance systems that inform the driver of situations on the road. This has been applied to two use cases. The first one, a variable message sign (VMS) detector, which takes images of the road, locates them, and announces their content. It is based on a deep learning model and a pipeline that processes the image, extracting the text using the optical character recognition model and reproducing the content using a cloud service. The second, an analyzer of circular Spanish roundabouts that, using aerial images, recognizes the different lanes and vehicles and extracts information about their status. This system combines several computer vision algorithms to recognize the circumferences of the lanes and a deep learning model that detects the different types of vehicles.