Identificación de regiones libres de obstáculos utilizando una red neuronal convolucional

Abstract

This research presents the configuration, train and evaluation of a convolutional neural network intended to

detect floor regions in point clouds. This is part of the development of an assisted navigation device for visually impaired people. In order that network’s processing time is adequate for this device, the identification of floor regions was tackled as a classification problem of patches generated by a sliding window. But unlike other solutions that use this approach, this work proposes labeling all the pixels of the patch and not overlapping the window with which they are extracted. The net is evaluated in public datasets Labelmefacade and NYU-v2 as well as in a private dataset of scenes retrieved from Faculty of Engineering Systems of Escuela Politécnica Nacional building. Tests conducted in public datasets show a reduction in processing time with respect of state-of-the-art models. Tests conducted in the private dataset show that the net reduces the effect of the noise generated during information retrieving and keeps its compactness, thus guaranteeing a efficient information processing.