Use this link to cite:
http://hdl.handle.net/2183/23739 Uso de técnicas de machine learning para realizar mapping en robótica móvil
Loading...
Identifiers
Publication date
Authors
Cebollada, Sergio
Román, Vicente
Payá, Luis
Tenza, María Flores
Jiménez, Luis M.
Reinoso, Óscar
Advisors
Other responsabilities
Journal Title
Bibliographic citation
Cebollada, S., Román, V., Payá, L., Tenza, M.F., Jiménez, L.M., Reinoso, O. (2019). Uso de técnicas de machine learning para realizar mapping en robótica móvil. En XL Jornadas de Automática: libro de actas, Ferrol, 4-6 de septiembre de 2019 (pp. 686-693). DOI capítulo: https://doi.org/10.17979/spudc.9788497497169.686. DOI libro: https://doi.org/10.17979/spudc.9788497497169
Type of academic work
Academic degree
Abstract
[Resumen] El trabajo presentado consiste en un estudio de diversos
clasi cadores basados en machine learning
como herramienta para llevar a cabo la tarea de
mapeo y localizaci on en rob otica m ovil. En concreto,
estos clasi cadores son utilizados para solventar
la tarea de localizaci on "gruesa", la cual forma
parte de los procesos a realizar para resolver la
localizaci on jer arquica. El proceso de localizaci on
llevado a cabo por el robot consistir a en (1) capturar
una imagen desde una posici on desconocida,
(2) calcular su correspondiente descriptor de apariencia
global, (3) introducir dicha informaci on al
clasi cador y obtener la estancia en la cual se encuentra
el robot en ese instante. Tras esto, (5) el
robot realizar a el problema de image retrieval con
toda la informaci on visual de entrenamiento contenida
en la estancia seleccionada (localizaci on -
na). Este trabajo eval ua el uso de tres clasi cadores
(Na ve Bayes, SVM y clasi cador basado en
red neuronal) los cuales se entrenan con tres posibles
descriptores de apariencia global (HOG, gist
y un descriptor obtenido a partir de una CNN).
Los experimentos se llevan a cabo mediante el uso
de un dataset que contiene im agenes omnidireccionales
capturadas en entornos de interior y que
presenta cambios din amicos (personas andando,
cambios de mobiliario, etc.). Los resultados obtenidos
demuestran que el m etodo propuesto es una
alternativa e ciente para realizar la tarea de localizaci
on jer arquica en cuanto error de localizaci on
y tiempo de c omputo
[Abstract] This work introduces a study regarding the use of several classi ers based on machine learning tools to carry out the mapping and localization task in mobile robotics. These classi ers are used to solve the rough localization, which is part of the hierachical localization process. Therefore, the localization tackled by the robot consists in (1) obtaining an image from an unknown position, (2) calculating its related global appearance descriptor, (3) puting this information into the classifer to estimate the current room. Afterwards, (5) the robot carries out the image retrieval problem with all the visual information provided by the training dataset contained in the selected room ( ne localization step). This work evaluates the use of three types of classi ers (Na ve Bayes, SVM and a classi er based on neural networks) which are trained with three possible global appearance descriptors (HOG, gist and a descriptor calculated from a CNN). The experiments are carried out through the use of a dataset which contains omnidirectional images captured indoor under dynamic changes (people walking, furniture changes, etc.). The results obtained show that this method proposed is an e cient alternative to tackle the hierarchical localization regarding the localization error and the computing time.
[Abstract] This work introduces a study regarding the use of several classi ers based on machine learning tools to carry out the mapping and localization task in mobile robotics. These classi ers are used to solve the rough localization, which is part of the hierachical localization process. Therefore, the localization tackled by the robot consists in (1) obtaining an image from an unknown position, (2) calculating its related global appearance descriptor, (3) puting this information into the classifer to estimate the current room. Afterwards, (5) the robot carries out the image retrieval problem with all the visual information provided by the training dataset contained in the selected room ( ne localization step). This work evaluates the use of three types of classi ers (Na ve Bayes, SVM and a classi er based on neural networks) which are trained with three possible global appearance descriptors (HOG, gist and a descriptor calculated from a CNN). The experiments are carried out through the use of a dataset which contains omnidirectional images captured indoor under dynamic changes (people walking, furniture changes, etc.). The results obtained show that this method proposed is an e cient alternative to tackle the hierarchical localization regarding the localization error and the computing time.
Description
Editor version
Rights
Atribución-NoComercial-CompartirIgual 4.0


