Categorização de ambientes internos a partir de imagens obtidas por câmeras acopladas a robôs terrestres móveis

Abstract

Over the last decade, there has been a significant increase in interest in incorporating semantic information into autonomous robot navigation. In this work, we propose an approach to map indoor scenes from images captured by a ground robot through Deep Learning. The methodology is based on an integrated representation learning model composed of three components: i) an autoencoder; ii) a You Only Look Once (YOLO) object detection model; and iii) a Convolutional Neural Network (CNN). The autoencoder and the CNN are employed for learning the representation from images of indoor scenes, while the YOLO is used for representing the semantic information. Experiments are conducted on two well-established image datasets for place categorization in indoor environments. The results demonstrate that our approach achieves significant accuracy in different scenarios, with dynamic objects, people, and lighting conditions. This high level of accuracy showcases the robustness and efficiency of our proposed method in the semantic mapping of indoor scenes.