A Bayesian framework for informed search using convolutions between observation likelihoods and spatial relation masks

In this work, a novel methodology for robots executing informed object search is proposed. The methodology is based mainly on a Bayesian framework that uses convolutions between observation likelihoods and spatial relation masks for estimating the probability map of the object being search for. By using spatial relation masks, complex spatial relations between objects can be defined as weighted sums of basic spatial relations using co-occurrence matrices as weights. The methodology is validated in an office environment in which four object classes (“monitor,” “keyboard,” “system unit,” and “router”) and four basic spatial relations (“very near,” “near,” “far,” and “very far”) are considered. Experiments combine statistics about object´s coocurrence and about object detection in real environments, and search trials using a realistic simulation tool in which extensive tests comparing six object search algorithms are carried out. Results show that the use of the proposed methodology increases the object detection rate in a search task from 27.5% up to 53.2%.