|Thesis abstract: |
My research work deals with the problem of building semantic models of indoor environments. A semantic model is a high-level form of representation that associates semantic labels to physical portions of the environments.
Classic approaches build semantic maps by labelling parts of the environments (e.g., rooms as `corridors¿ or `offices¿) only relying on sensorial data, that capture their physical and geometrical features.
We propose a novel building-level approach that exploits the a priori knowledge of the building typology, using some results from the field of architecture as a source. In a nutshell, our approach extracts relevant features and statistics from floor plans of actual buildings using data mining techniques; the resulting knowledge base is then processed with machine learning algorithms in order to create generative models of semantic maps for specific types of buildings, which can be used, integrated with the data coming from sensors, to predict the structure of unobserved parts of a building.
The main target application of our models are autonomous mobile robots: the goal is to give them the ability to discover and predict an unknown environment using prior knowledge of other similar structures, mimicking human behaviours used for navigating.
Specifically, our semantic models are employed to inform the exploration of unknown environments and to improve quality and realism of 3D simulations used in robotics, which represent powerful tools to test the behaviour of robots in a cheap and repeatable fashion.