Place Classification using Dempster-Shafer Theory

The paper presents a novel place labeling method. It is assumed that an indoor mobile robot equipped with a camera or RGB-D sensor ambulates an indoor environment. The places visited by the robot are classified based on objects which have been recognized. Each object (or set of objects) votes for a set of room classes. Data aggregation is performed using Dempster-Shafer theory (DST), which can be regarded as a generalization of the Bayesian theory. The possibility of taking into account the uncertainty of data is the main advantage of the described method. The classic Dempster’s rule of data aggregation has been criticized because it can lead to non-intuitive results. Many alternative methods have been proposed and several were tested during our experiments. Most place classification methods assume a closed world model, i.e. a test sample is assigned to the most probable class even if its corresponding probability is very small. An advantage of our system is the intrinsic capability of giving unknown class as an answer in such situations, which can be used by the robot to take appropriate actions.