Passive Presence Detection is the task of detecting when a person is present or not, without the need for them to carry any wireless device. For that, a wireless sensor network consisting of Bluetooth Low Energy (BLE) devices, often called nodes or beacons must be employed in an area of interest in which their positions are not considered known. Beacons transmit and receive BLE signals and they report the Received Signal Strength (RSS) of them, counted in decibels. The RSS data can be explored in search of patterns that can classify presence versus non-presence of people near a wireless sensor network. Machine Learning and Signal Processing techniques are used to tackle the classification problem of Passive Presence Detection. Specifically, Random Forests and two versions of a Sparse Representationbased classification model are employed. Random Forests create an ensemble of classification trees, based on random features. Each of the trees votes for a class. The Sparse representation-based classifier exploits the theory of compressed sensing to find sparse representations of the RSS data, i.e new data matrices which contain only a few non-zero elements. Compressed sensing states that if a data set has a sparse representation in some overcomplete basis, then a high quality reconstruction is possible with much fewer measurements than we would normally need. This method finds an overcomplete basis for each class and classifies the input data based on the most accurate reconstruction. Two data sets were used under various experimental settings to verify the efficacy of the proposed methods. Both Random Forests and the Sparse Representation-based classifier report 100% classification accuracy.

Velden, M. van de
hdl.handle.net/2105/39577
Econometrie
Erasmus School of Economics

Sofos, A. (Andreas). (2017, October 5). Evaluation of Machine Learning techniques for Passive Presence Detection. Econometrie. Retrieved from http://hdl.handle.net/2105/39577