A bayesian based methodology for indirect object search

Abstract

The main goal of this paper is to propose a Bayesian based methodology for implementing robot informed search for objects. The methodology uses convolutions between observation likelihoods of secondary objects and spatial relation masks for estimating the probability map of the object being searched for, and also a search procedure that uses this probability map. A method for computing complex spatial relation masks by using a basis composed of basic relation masks and a database of co-occurrences of objects is used. Each basic relation mask corresponds to a qualitative spatial relation (QSR), such as: 'very near', 'near', or 'far'. The search procedure takes into account the probability that the main object can be in different regions on the map and the distance to those regions. Also, the object search procedure is able to detect objects and generate new plans while moving. The proposed methodology is compared with uninformed and alternative informed search approaches using simulations and real-world experiments with a service robot. In simulations, the use of the proposed methodology increases the detection rate from 28% (direct uninformed search) to 79%, when the main object can be detected within a maximum distance of 1 meter. In the real world experiments, the use of the proposed methodology increases the detection rate from 40% (direct uninformed search) to 87% when using convolutions with soft masks, global search, and information on the positive detection of secondary objects. The detection rates obtained when using the proposed methodology are also much higher than those obtained by alternative informed search methods, both in the simulated and in the real-world experiments.