Sistemas de visión para el seguimiento de poses 3-d de objetos en tiempo real

• Autores: Leonardo Rubio Navarro
• Directores de la Tesis: Eduardo Ros Vidal (dir. tes.), Mancia Anguita López (dir. tes.), Javier Díaz Alonso (dir. tes.)
• Lectura: En la Universidad de Granada ( España ) en 2014
• Idioma: español
• Tribunal Calificador de la Tesis: Bernabe Linares Barranco (presid.), Samuel F. Romero García (secret.), Fernando Pardo Carpio (voc.), Rahul Swaminathan (voc.), Eva Martínez Ortigosa (voc.)
• Enlaces
  • Tesis en acceso abierto en: DIGIBUG
• Resumen

  • This dissertation presents a work related with image analysis and 3D object pose estimation. A wide variety of object's types and pose estimation methods have been addressed along this PhD manuscript. Multiple pose estimation systems have been implemented with improvements that outperform state-of-the-art methods. The comparison is carried out in real-world and synthetic benchmarks where the ground-truth is known and error measurement mechanisms are provided.

    This dissertation is structured in four main parts: In the first part, an overview of the state-of-the-art is exposed. We review the different camera representations used in computer vision systems. We continue with a description of the different computer vision features and the algorithms for their extraction. In the final part of this section we define the pose estimation algorithms that use the camera's representation and the extracted visual features to compute the object pose estimation.

    After the literature's overview, we continue with a second part that states a generic architecture for object pose estimation. This architecture defines a development environment for further contributions. We develop a detection pose estimation system that use single-frame information and is based on the state-of-the-art methods. We combine this system with tracking methods that use temporal information in order to implement a robust and accurate pose estimation system. A comparison with the literature's systems is carried out using a synthetic benchmark that have been produced for this comparative study.

    Following the generic architecture defined in the previous part, this third section explains one of the main contribution related with articulated object detection. A detection system for articulated object is described where features, models and algorithms are adapted for this purpose. A synthetic benchmark for articulated objects was developed along with error measurement methods that allow the system comparison with the state-of-the-art algorithms.

    Another main contribution of this work is related with the improvements adopted in pose estimation systems that aim an enhanced performance in terms of accuracy, execution time and robustness. The improvements involve object's model simplification and environmental knowledge adaptation to the pose estimation process. Using augmented reality systems and depth cameras, we create a benchmark for real-world rigid object tracking that allows the comparison with alternative methods. We also extend the system applications to diverse situations such as camera pose estimation.

    As conclusions, this dissertation implements multiple pose estimation systems for different types of objects in real-world scenarios. Proposed systems are compared with alternative methods through the developed benchmarks. The different methods evaluation aims remarkable results.