Seguimiento de múltiples objetos en entornos interiores muy poblados basado en la combinación de métodos probabilísticos y determinísticos

  1. Marrón Romera, Marta
Supervised by:
  1. Juan Carlos García García Director
  2. Miguel Angel Sotelo Vázquez Co-director

Defence university: Universidad de Alcalá

Fecha de defensa: 07 October 2008

  1. Felipe Espinosa Zapata Chair
  2. María Elena López Guillén Secretary
  3. José María Cañas Plaza Committee member
  4. Arturo de la Escalera Hueso Committee member
  5. Daniel Gatica Pérez Committee member
  1. Automática

Type: Thesis


This thesis is involved in the area of service and personal robotics, a research topic that has gained importance in the last two decades thanks to the evolution of technology and its insertion in everyday life of people. In this context, the thesis proposes a multiple target tracking (MTT) algorithm appropriate to be used in complex indoor environments. The MTT gives information about number, position, speed, track and identification of the different objects in the environment of the robot, information extracted from the data collected by the input observation system. The proposed solution fulfils some specifications derived from the desired performance: it has to consider the uncertainty of objects behaviour and sensor models; it has to be flexible in order to process input data coming from different kinds of sensors (vision, sonar, audio, laser, radio, etc., or some kind of fusion among any of them); it must handle different types of objects that can affect the robot’s behaviour in its environment, despite of their dynamics and deformable shape; it has to accomplish real time execution regardless of the number of objects and complexity of the scene being tracked; finally, it has to achieve the level of robustness and reliability needed by personal and service robotics, where safety of the surrounding objects (persons and other robots) and the robot itself is a main requirement. In order to achieve all the fore mentioned specifications, a combination of probabilistic and deterministic algorithms is proposed as the best solution for the MTT. A particle filter is used as estimation kernel of the tracker, and two deterministic clustering proposals are incorporated to it, where they are used as association process and output filtering, respectively, in the tracking task. This combination results in the “Extended Particle Filter with Clustering Process”, XPFCP, the proposal for multi-tracking applications presented in this thesis. Particle filters have the capability to model multiple states within a single distribution and have a great flexibility in managing any kind of dynamics and observation models. These characteristics make this version of the Bayes filter the most suitable algorithm for the multitracking solution proposed. Therefore, they can accomplish the tracking task pursuit with an almost constant computational cost. The same approach has been tried in some other research works, but the lack of robustness of the final MTT implementation has lead to discarding these solutions. This thesis proposes and demonstrates that a deterministic part in the MTT adds the robustness that the multimodal estimator needs. Throughout this thesis it can be found a deep revision of previous works (algorithms and results) carried out by the scientific community in the topic of interest. Then, it has been made an exhaustive study of the proposed tracker behaviour in complex tracking tasks in terms of reliability, efficiency, robustness and execution time. Finally, a comparison among the proposed solution and two of the best known and most widely used multi- tracking algorithms in the scientific community (the “Joint Probabilistic Data Association Filter” in its continuous –JPDAF– and sampled –SJPDAF– versions) is also performed and analyzed in order to validate the thesis contribution in this research area.