Searching and tracking of humans in urban environments with humanoid robots
- Autores: Alex Goldhoorn
- Directores de la Tesis: Alberto Sanfeliu Cortés (dir. tes.), René Alquézar Mancho (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2017
- Idioma: español
- Tribunal Calificador de la Tesis: Juan Andrade Cetto (presid.), Vicente Matellán Olivera (secret.), Luis Merino (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
Searching and tracking are important behaviours for a mobile service robot to assist people, to search-and-rescue and, in general, to locate mobile objects, animals or humans. Even though searching might be evident for humans, for robots it is not, since it requires exploring, handling noisy sensors, coping with dynamic obstacles, and coordination in the case of multiple agents.
In this thesis, we present several methods to search and track a person in an urban environment. All methods were fist tested extensively in simulation and then in real-life, using one or two mobile service robots, called Tibi and Dabo. The robots have laser range finders, which are used to navigate, to detect obstacles and to detect people's legs. Since we focus on search-and-track methods, we use existing methods for robot navigation, for people detection and person recognition.
First tests are done with the hide-and-seek problem, in which the robot learns to catch the hider. Concretely, a Mixed Observable Markov Decision Process (MOMDP) model is used, in which the seeker's location is fully observable and the hider's location partially observable. Since the computational complexity depends on the number of states, we propose a hierarchical on-line method that reduces the state space by grouping them together. Although the method worked properly in simulation, in the real-life experiments the results were not satisfying and the on-line policy calculation was not fast enough to work in real-time.
To handle larger environments, work in continuous state space and run in real-time, we propose to use an approach, the Continuous Real-time POMCP (CR-POMCP), that does Monte-Carlo simulations to learn a policy. The method performed correctly in simulation, but on the real robot it resulted in slow zigzag movements. Therefore, a new method is proposed, which uses the highest probable locations, according to its probability map (belief). Since the belief propagation of the POMCP resembles how a Particle Filter (PF) works, we also propose a method that uses a PF to maintain the belief. The PF method has to handle lack of observations, therefore, we introduce a special weight function. Both belief update methods take into account sensor and actuator noise, false negative detections, false positive detections (for a short time) and dynamic obstacles.
Finally, a cooperative distributed multi-agent method is presented, it makes use of the previous belief update functions, but it uses all the agents' observations. Next, the search locations are assigned by exploring the working environment, taking into account: the belief, the distance to the search location and if another agent already will search close to it. Summarizing, the main contributions of this thesis are several methods to search and track a person in an urban environment with one or more mobile service robots. All these methods have been shown to work through a set of simulations and real-life experiments.
