Neural network-based robot localization using visual features

Abstract

This paper outlines the development of a module capable of constructing a map-building algorithm using inertial odometry and visual features. It incorporates an object recognition module that leverages local features and unsupervised artificial neural networks to identify non-dynamic elements in a room and assign them positions. The map is modeled using a neural network, where each neuron corresponds to an absolute position in the room. Once the map is constructed, capturing just a couple of images of the environment is sufficient to estimate the robot's location. The experiments were conducted using both simulation and a real robot. The Webots environment with the virtual humanoid robot NAO was used for the simulations. Concurrently, results were obtained using a real NAO robot in a setting with various objects. The results demonstrate notable precision in localization within the two-dimensional maps, achieving an accuracy of ± (0.06, 0.1) m in simulations contrasted with the natural environment, where the best value achieved was ± (0.25, 0.16) m.