Monocular slam: data association and sensing through a human-assisted uncalibrated visual system
- Autores: Edmundo Guerra Paradas
- Directores de la Tesis: Antoni Grau Saldes (dir. tes.), Rodrigo Francisco Munguía Alcalá (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2017
- Idioma: inglés
- Tribunal Calificador de la Tesis: Alberto Sanfeliu Cortés (presid.), Jorge Palacín Roca (secret.), Lluís Ribas Xirgo (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
The Simultaneous Localization and Mapping (SLAM) problem is widely acknowledged as one of the fundamental problem to solve in perception and robotics to produce actual mobile robotic agents. The problem itself is that of how can a mobile robot agent operate in an a priori unknown environment, using the sensory systems available (normally on itself) to perceive its surroundings, build a map with this knowledge, and localize itself in said map tracking its own position.
This relevance, combined with the diversity of approaches available to solve it, and the depth of the challenges it presents, makes the SLAM problem one of the more active areas of research in robotics. One of the most complex challenges in any approach is the data association, as it generally conveys hard a trade-off between robustness and computational time required, and can impact the whole architecture of a SLAM method.
In terms of sensors used, the field was originally dominated by range finder sensors, but visual SLAM research has grown in popularity in the last decade. Camera sensors have been expanding its capabilities and specifications thanks to the consumer demand for them. As a sensor, they provide lightning measurements of the projected points at known bearings, which through computer vision can be converted into bearing measurements for visual features, which can be themselves of several levels of complexity.
The same consumer demand has also pushed technical developments in MEMS and robotic devices with a direct impact in the field of cooperative robotics and the emergence of wearable device technology, where human can wear or carry devices with several sensors in an unobtrusive way. These technologies have opened many opportunities in for research in robotics, including the field of collaborative SLAM and the area of human-robot interaction (HRI).
This thesis is focused in the study and development of a visual SLAM methodology based on the delayed inverse-depth feature initialization (DI-D) monocular SLAM which can benefit and exploit the advantages of working in a HRI collaborative framework. In order to achieve this, the research is focused in two different areas. Firstly, the known and tested DI-D monocular SLAM is studied: its procedures and algorithms detailed and analyzed; with emphasis in the data association problem (DA). The DA process is reviewed, and a new validation algorithm is introduced to strengthen and give robustness to the data association technique used.
Once the DI-D has been studied and updated the HRI collaborative framework is introduced, with an initially focus into solving one of its inconveniences: the requirement of a scaled metric initialization with a priori knowledge. The HRI is introduced by deploying into a human being a custom built wearable device which includes a camera and some other sensors. The data from this secondary monocular sensor, whose pose is approximately known with respect to the camera used to solve the SLAM problem, allows speeding up the feature initialization process of the DI-D, and even ignoring the requirement of scale initialization.
As the introduction of the HRI framework was successful, its advantages were further expanded to the rest of the SLAM process, including the measurement and update steps. This integration was performed based in a virtual sensor methodology, where the collaborative measurement process was treated as a single sensor with its specifications, allowing seamless fusion into the EKF-SLAM (Extended Kalman Filter SLAM). To evaluate the specific impact of the HRI with respect to the behaviour of the secondary camera, several new metrics have been proposed and studied.
All the methods have been proved and validated through experimentation with real data. When it was found relevant, the experiments were evaluated in real-time scenarios, and several simulations have been included when needed to prove some theoretical hypothesis.
