Scene understanding for autonomous robots operating in indoor environments

• Autores: Alejandra Carolina Hernández Silva
• Directores de la Tesis: Ramón I. Barber Castaño (dir. tes.), Óscar Martínez Mozos (codir. tes.)
• Lectura: En la Universidad Carlos III de Madrid ( España ) en 2021
• Idioma: español
• Tribunal Calificador de la Tesis: Carlos Balaguer Bernaldo de Quirós (presid.), Fernando Matía Espada (secret.), Klaus Strobl (voc.)
• Programa de doctorado: Programa de Doctorado en Ingeniería Eléctrica, Electrónica y Automática por la Universidad Carlos III de Madrid
• Enlaces
  • Tesis en acceso abierto en: e-Archivo
• Resumen

  • The idea of having robots among us is not new. Great efforts are continually made to replicate human intelligence, with the vision of having robots performing different activities, including hazardous, repetitive, and tedious tasks. Research has demonstrated that robots are good at many tasks that are hard for us, mainly in terms of precision, efficiency, and speed. However, there are some tasks that humans do without much effort that are challenging for robots. Especially robots in domestic environments are far from satisfactorily fulfilling some tasks, mainly because these environments are unstructured, cluttered, and with a variety of environmental conditions to control.

    This thesis addresses the problem of scene understanding in the context of autonomous robots operating in everyday human environments. Furthermore, this thesis is developed under the HEROITEA research project that aims to develop a robot system to help elderly people in domestic environments as an assistant. Our main objective is to develop different methods that allow robots to acquire more information from the environment to progressively build knowledge that allows them to improve the performance on high-level robotic tasks. In this way, scene understanding is a broad research topic, and it is considered a complex task due to the multiple sub-tasks that are involved. In that context, in this thesis, we focus on three sub-tasks: object detection, scene recognition, and semantic segmentation of the environment.

    Firstly, we implement methods to recognize objects considering real indoor environments. We applied machine learning techniques incorporating uncertainties and more modern techniques based on deep learning. Besides, apart from detecting objects, it is essential to comprehend the scene where they can occur. For this reason, we proposed an approach for scene recognition that considers the influence of the detected objects in the prediction process. We demonstrate that the exiting objects and their relationships can improve the inference about the scene class. We also consider that a scene recognition model can benefit from the advantages of other models. We propose a multi-classifier model for scene recognition based on weighted voting schemes. The experiments carried out in real-world indoor environments demonstrate that the adequate combination of independent classifiers allows obtaining a more robust and precise model for scene recognition.

    Moreover, to increase the understanding of a robot about its surroundings, we propose a new division of the environment based on regions to build a useful representation of the environment. Object and scene information is integrated into a probabilistic fashion generating a semantic map of the environment containing meaningful regions within each room. The proposed system has been assessed on simulated and real-world domestic scenarios, demonstrating its ability to generate consistent environment representations.

    Lastly, full knowledge of the environment can enhance more complex robotic tasks; that is why in this thesis, we try to study how a complete knowledge of the environment influences the robot's performance in high-level tasks. To do so, we select an essential task, which is searching for objects. This mundane task can be considered a precondition to perform many complex robotic tasks such as fetching and carrying, manipulation, user requirements, among others. The execution of these activities by service robots needs full knowledge of the environment to perform each task efficiently. In this thesis, we propose two searching strategies that consider prior information, semantic representation of the environment, and the relationships between known objects and the type of scene. All our developments are evaluated in simulated and real-world environments, integrated with other systems, and operating in real platforms, demonstrating their feasibility to implement in real scenarios, and in some cases outperforming other approaches. We also demonstrate how our representation of the environment can boost the performance of more complex robotic tasks compared to more standard environmental representations.