Indoor Robot Localization in Hand-Drawn Maps by using Convolutional Neural Networks and Monte Carlo Method

Abstract

Localization for estimating the position of a robot in an environment remains a challenging problem in mobile robots. However, previous studies mostly consider accurate map with correct scale before the localization task. This paper presents a novel approach for solving localization problem using the inaccurate hand-drawn map of the environment, when the exact map of the environment is not prepared before the localization task. Our proposed method firstly decomposes the hand-drawn map into the local places such as a room or corridor, then extract a set of geometric information of each segmented area to train them using convolutional neural networks (CNN) for place recognition. This technique only selects nominated segmented areas of where the possible location of the robot is. Secondly, Monte Carlo Localization (MCL) technique is used to estimate the position of the robot. Empirical studies on the standard localization technique illustrate that the proposed approach achieves superior performance to state-of-the-art localization methods regarding noisy data issues and large localization error.