Reconstruction and recognition of confusable models using three-dimensional perception

• Autores: Jorge García Bueno
• Directores de la Tesis: Luis Enrique Moreno Lorente (dir. tes.)
• Lectura: En la Universidad Carlos III de Madrid ( España ) en 2013
• Idioma: inglés
• Tribunal Calificador de la Tesis: Alfonso José García Cerezo (presid.), María Dolores Blanco Rojas (secret.), Carlos Sagüés (voc.)
• Materias:
  • Matemáticas
    • Ciencia de los ordenadores
      • Inteligencia artificial
• Enlaces
  • Tesis en acceso abierto en: e-Archivo
• Resumen

  • Perception is one of the key topics in robotics research. It is about the processing of external sensor data and its interpretation. The necessity of fully autonomous robots makes it crucial to help them to perform tasks more reliably, flexibly, and efficiently. As these platforms obtain more refined manipulation capabilities, they also require expressive and comprehensive environment models: for manipulation and affordance purposes, their models have to involve each one of the objects present in the world, coincidentally with their location, pose, shape and other aspects. The aim of this dissertation is to provide a solution to several of these challenges that arise when meeting the object grasping problem, with the aim of improving the autonomy of the mobile manipulator robot MANFRED-2. By the analysis and interpretation of 3D perception, this thesis covers in the first place the localization of supporting planes in the scenario. As the environment will contain many other things apart from the planar surface, the problem within cluttered scenarios has been solved by means of Differential Evolution, which is a particlebased evolutionary algorithm that evolves in time to the solution that yields the cost function lowest value. Since the final purpose of this thesis is to provide with valuable information for grasping applications, a complete model reconstructor has been developed. The proposed method holdsmany features such as robustness against abrupt rotations, multi-dimensional optimization, feature extensibility, compatible with other scan matching techniques, management of uncertain information and an initialization process to reduce convergence timings. It has been designed using a evolutionarybased scan matching optimizer that takes into account surface features of the object, global form and also texture and color information. The last tackled challenge regards the recognition problem. In order to procure with worthy information about the environment to the robot, a meta classifier that discerns efficiently the observed objects has been implemented. It is capable of distinguishing between confusable objects, such as mugs or dishes with similar shapes but different size or color. The contributions presented in this thesis have been fully implemented and empirically evaluated in the platform. A continuous grasping pipeline covering from perception to grasp planning including visual object recognition for confusable objects has been developed. For that purpose, an indoor environment with several objects on a table is presented in the nearby of the robot. Items are recognized from a database and, if one is chosen, the robot will calculate how to grasp it taking into account the kinematic restrictions associated to the anthropomorphic hand and the 3D model for this particular object. -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------