Entrenamiento, optimización y validación de una CNN para localización jerárquica mediante imágenes omnidireccionales
- Juan José Cabrera 1
- Sergio Cebollada 1
- Mónica Ballesta 1
- Luis Miguel Jiménez 1
- Luis Payá 1
- Óscar Reinoso 1
-
1
Universidad Miguel Hernández de Elche
info
Publisher: Universitat Jaume I ; Servizo de Publicacións ; Universidade da Coruña ; Comité Español de Automática
ISBN: 978-84-9749-804-3
Year of publication: 2021
Pages: 640-647
Congress: Jornadas de Automática (42. 2021. Castellón)
Type: Conference paper
Abstract
The aim of this work is to address the localization of a mobile robot by training a Convolutional Neural Network (CNN) in order to obtain optimal results. The localization problem is approached in a hierarchical way by using an omnidirectional catadioptric system and working directly with the captured images without panoramic conversion, saving the computational time associated with this process. Localization is carried out in two steps, both using the CNN architecture for different purposes. First, a rough localization is carried out, which consists of identifying the room in which the robot is located by means of the CNN. Then a fine localization is performed in the room, in which the CNN is used to obtain holistic descriptors from the intermediate layers of the network. These global-appearance descriptors allow finding the position where the robot is located more precisely by means of a nearest neighbour search, comparing the corresponding descriptor of the test image with the descriptors of the images captured in the room selected in the first step. In order to improve the accuracy of the network, data augmentation and Bayesian hyperparameter optimisation are used. These techniques prove to be an efficient and robust solution to tackle the localization problem as shown in the experiments section.