Semantic Localization through Propagation of Scene Information in a Hierarchical Model

The success of mobile robots, and particularly these coexisting with humans, relies on the ability to understand human environments. Representing the world and analysing spaces in a similar way to humans will enhance their comprehension and enable higher abstraction capabilities and interactions. The purpose of this work is to develop a localization framework that takes into account the different scenes common in a human environment and a hierarchical model of the environment. A probabilistic model for recognizing scenes is employed to determine the scene in which the robot is located. To allow that, the information about the objects and the relationships between them are considered. Besides that, a hierarchical model formed by different topological representations according to different levels of abstraction is proposed. Localization is performed at different levels to improve the localization accuracy. In this work, scene information is used to improve the localization of a mobile robot in a hierarchical model using hidden Markov models. The experiments of our framework working in real environments uphold the usefulness of the inclusion of the understanding and abstraction of the environment in localization.