Ensemble Learning for Seated People Counting using WiFi Signals: Performance Study and Transferability Assessment

The detection, location, and behavior recognition of human beings in different environments is not only a subject of a wide range of studies, but has also triggered the development of a large number of applications, including those which enhance sustainability and efficiency of infrastructures. For instance, the estimation of the occupancy could improve the energy management of a building. Due to human presence or movement over a particular area, the analysis of variations in wireless signal properties of already deployed wireless technology such as WiFi systems provides the information needed for Machine Learning models to accomplish the non-intrusive (device-free) detection and classification of different human activities. In this context, this work focuses on detecting seated people in an indoor scenario by using ensemble learning, a particular branch of Machine Learning models for supervised learning that hinges on combining the outputs of individual predictors. Furthermore, we evaluate the transferability of the knowledge modeled by ensemble learners. When trained in a particular frequency or channel, such models are used to classify data captured over another different frequency. Our experimental setup and discussed results reveal that while ensembles attain satisfactory levels of predictive accuracy predictions, their knowledge cannot be transferred among different frequencies. This conclusion opens an exciting future towards new means to perform effective knowledge transfer over the frequency domain.