Detecting Latent Variables of Interest in Geo-Localized Environments Using an Aerial Robot

In general, monitoring applications require human intervention whenever there is no physical sensors for the variables of interest (e.g. People in danger after a catastrophe). In this paper we describe an inference engine which is used to estimate latent variables that can not be perceived by sampling the physical phenomena directly. Our approach uses information from different types of sensors, and fuses them along with knowledge of experts. The inference engine works with probabilistic first order logic rules based on geo-located sensed data as evidences in order to dynamically create the structure of a Bayesian network. Our experiments, performed by using an aerial robot with a mounted RGB-Camera, show the capability of our method to detect people in danger situations, where the physical variables to being sensed are humans and fire.