Detecció de punts de referència en entorns urbans mitjançant visió per computador

Abstract

[en] This project aims to supply the ADD autonomous robot with a landmark detection system that should be able to easily integrate with the existing localization stack. Multiple landmark have been developed, but this project is centered on urban traffic signs. To be able to detect and classify the signs, three different methods are proposed: the first uses a convolutional neural network (CNN) to both detect and classify the signs; the second one uses a CNN to detect and another one to classify; the third, uses a CNN to detect and more traditional methods to classify. Multiple networks and optimizations are considered, with a strong emphasis with those on the YOLO family. These methods have been tested with the well known GTSDB and GTSRB, and some results have been comparable to the ones obtained in the state of the art. They are also tested in environments much more similar to the ones the robot will encounter, using a specially made dataset.