Estimating occupancy in heterogeneous sensor environment

A general approach is proposed to determine the common sensors that shall be used to estimate and classify the approximate number of people (within a range) in a room. The range is dynamic and depends on the maximum occupancy met in a training data set for instance. Means to estimate occupancy include motion detection, power consumption, CO2 concentration sensors, microphone or door/window positions. The proposed approach is inspired by machine learning. It starts by determining the most useful measurements in calculating information gains. Then, estimation algorithms are proposed: they rely on decision tree learning algorithms because these yield decision rules readable by humans, which correspond to nested if-then-else rules, where thresholds can be adjusted depending on the living areas considered. In addition, the decision tree depth is limited in order to simplify the analysis of the tree rules. Finally, an economic analysis is carried out to evaluate the cost and the most relevant sensor sets, with cost and accuracy comparison for the estimation of occupancy. C45 and random forest algorithms have been applied to an office setting, with average estimation error of 0.19-0.18. Over-fitting issues and best sensor sets are discussed.