A Localization Approach Based on Omnidirectional Vision and Deep Learning
S. Cebollada, L. Payá, M. Flores, V. Román, A. Peidró, O. Reinoso
Lecture Notes in Electrical Engineering. Informatics in Control, Automation and Robotics (2022)
Ed. Springer ISBN:978-3-030-92441-6 ISSN:1876-1100 DOI:https://doi.org/10.1007/978-3-030-92442-3_13 - 793 (226-246)
The present work introduces a study about the use of a deep learning tool to tackle the visual localization. The approach proposed consists in developing a Convolutional Neural Network (CNN) with the aim of addressing the room retrieval task. Additionally, the network can be used to extract holistic descriptors from intermediate layers. Therefore, the localization can be carried out by meas of comparing the holistic descriptor obtained during the localization process with the descriptors obtained during the mapping process, but it can be carried out by using a hierarchical strategy. Concerning the hierarchical localization approach, in previous works, it has been addressed by means of a nearest neighbour search using different layers of information. In the present work, first is addressed a rough step which consists in solving the room retrieval with the CNN and, after that, a nearest neighbour search is carried out by using the holistic descriptors contained in the room selected. Hence, this work evaluates firstly the validity of the holistic descriptors extracted from the CNN and, secondly, evaluates the hierarchical method based on the CNN tool. The experiments to evaluate the validity of the proposed methods are carried out with an indoor dataset with real-operation conditions. The results show that the proposed approach based on deep learning is a robust solution to tackle the visual localization task.