Tracking and intercepting pedestrians: a robotic approach to surveillance of critical infrastructures
- Autores: Mario Andrei Garzón Oviedo
- Directores de la Tesis: Antonio Barrientos Cruz (dir. tes.), Jaime del Cerro Giner (codir. tes.)
- Lectura: En la Universidad Politécnica de Madrid ( España ) en 2016
- Idioma: inglés
- Tribunal Calificador de la Tesis: Rafael Aracil Santonja (presid.), Fernando Matía Espada (secret.), M. A. Olivares-Mendez (voc.), Miguel Ángel Salichs Sánchez-Caballero (voc.), Antonio Gimenez Fernandez (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: Archivo Digital UPM
- Resumen
This dissertation presents an approach for the inclusion of technologies from the field of robotics into security and surveillance of outdoor critical infrastructures. Nowadays, the systems used for these tasks are mainly reactive, meaning that they rely on alarms based on static sensors that require constant monitoring and a response from human operators. Therefore, the use of robotic systems can improve the capabilities of detecting and reacting to disruptions while reducing the risk for human operators. The global objective of this work is to develop a system capable of performing a high level task in the context of security systems: the detection, tracking and interception of an intruder moving around, in an outdoors critical infrastructure. Moreover, the work should be focused not only on developing algorithms and strategies but also on testing them with real robots in realistic scenarios. During the Thesis, a modular system has been developed to carry out the global mission, thus, the process has been divided into several modules, namely: pedestrian detection, tracking, trajectory prediction, planning for interception and autonomous navigation. This modular design allowed developing and testing the modules independently from each other. Moreover, it also allowed us to use them in combination with external algorithms or to apply them to different applications. Furthermore, in order to facilitate the integration of the modules and real experimentation, software and communications architectures have been developed. The pedestrians detection module is based on the fusion of information from two different sources: a 2D laser scanner and a camera. This module introduces two main novelties: the first one is the inclusion of an adaptive technique for projection of the region-of-interest, which addresses the problem of vertically localizing a projection from the laser into an image plane. The second one consists on introducing of a probability calibration step, applied before performing the fusion of information provided by both sensors, this allows reducing redundancy and correlation between two information sources. The next module included in this work is the pedestrian tracking. It keeps updated a list of detected pedestrians, including their position and velocities. In order to do that, the algorithm processes the observations and determines whether or not each observation corresponds to a previously detected pedestrian. The main novelty on this algorithm is the introduction of a discriminative tracking procedure, where the uncertainty of the detection is taken into account. This means that, highly certain detections are used to confirm the presence of a human in the scene, meanwhile uncertain ones are used to keep track of its position over large areas. The trajectory prediction module uses a comparison of the observed trajectory against possible routes in combination with a simple modelling of the movements, to obtain a long-term prediction of the future trajectory of the pedestrian. Its main contribution is that it does not require any previous observation of pedestrian trajectories to obtain the prediction. Moreover, it can be adapted to any scenario since it only requires a cost-map of the infrastructure environment. Also, the integration of a motion model allows estimating a long term time-stamped prediction, which is required by the interception task. The path planning for interception module constructs a tree of possible trajectories and evaluates each one of them in terms of their probability of intercepting the intruder according to its predicted trajectory. Subsequently, it extracts a route that likely intercepts a pedestrian, while avoiding both static and dynamic obstacles. The contribution in this case consisted on the modification and adaptation for this task, of an algorithm previously used for navigation in dynamic environments under uncertainty. A very important characteristic of the proposed system is the feasibility of being continuously updated as it is being executed. This means that any change in the movements of the pedestrian will be processed on-line and therefore, they will result in a modification of the trajectory of the pursuer robot. This makes it possible to obtain a system capable of performing a high level task, such as tracking and intercepting an intruder. No previous references have been found in literature presenting a real system performing that tasks in outdoor environments. The system has been intensively tested and characterized by using simulation scenarios, and it has also being tested on three real world scenarios, with different sizes and obstacle distributions. Thus allowing to demonstrate the correct integration of all sub-modules of the thesis, as well as the capabilities of the detection, tracking and interception system as a whole.
