The advent of Wi-Fi infrastructures and wireless devices has changed the wireless landscape considerably. Of particular interest is Wi-Fi metadata, which provides information about the devices present in a wireless network and offers unintrusive ways in the inference of wireless device location and behavior. The Wi-Fi metadata of mobile devices can be obtained by deploying a network of monitoring devices over Wi-Fi. Monitoring devices consist of network interface cards in monitoring mode and are capable of listening to all Wi-Fi packets in the network traffic. This thesis examines the problem of estimating the occupancy of a region based on Wi-Fi metadata and compares the performance obtained with the data gathered by monitoring system with directional antennas to that of the same system with isotropic antennas. The monitoring antenna design is important because certain antenna radiation patterns can yield more discriminating information about current conditions.
In this work, occupancy estimation is formulated as a classification problem, and logistic regression and radial basis function networks are employed as the classification schemes. The metadata from Wi-Fi packets combined with received signal strengths from the monitoring antennas act as input to the classifiers. This work considers a specific scenario of occupancy estimation. The classification performance of the system is assessed with the synthetic data at different noise levels in the RSSI values, and with the experimental data obtained from a testbed implementation. The field data are acquired in a line-of-sight environment. The findings indicate that the judiciously selected directional monitoring antennas work appropriately and significantly outperform the isotropic antennas in the scenario considered. This work introduces machine learning approaches for occupancy estimation based on Wi-Fi metadata and provides insight into the efficiency of directional antennas over isotropic antennas in the context of occupancy estimation. It also suggests that utilizing custom antenna designs based on the inference, shape and size of the target region can significantly improve the performance of a Wi-Fi monitoring system.
