Advancements in Wi-Fi-Based Passenger Counting and Crowd Monitoring: Techniques and Applications

The widespread use of personal mobile devices, including tablets and smartphones, created new opportunities for collecting comprehensive data on individual movements within cities while preserving their anonymity. Extensive research focused on turning personal mobile devices into tools for measuring human presence. To protect privacy, the data collected must be anonymous or pseudo-anonymous, leading to the preference for management data. A common approach involves analysing probe requests, which are Wi-Fi protocol messages transmitted by mobile devices while searching for access points. These messages contain media access control (MAC) addresses, which used to be unique identifiers. To safeguard the privacy of smartphone users, the major manufacturers (Google, Apple, and Microsoft) have implemented algorithms that generate random MAC addresses, which change often and unpredictably. This thesis focuses on the problem of fingerprinting Wi-Fi devices based on analysing management messages to overcome previous methods that relied on the MAC address and became obsolete. Detecting messages from the same source allows counting the devices in an area, calculating their permanence, and approximating these metrics with the ones of the humans carrying them. An open dataset of probe requests with labelled data has been designed, built, and used to validate the experiments. The dataset is also provided with guidelines for collecting new data and extending it. Since the dataset contains records of individual devices, the first step of this study was simulating the presence of multiple devices by aggregating multiple records in sets. Many experiments have been conducted to enhance the accuracy of the clustering. The proposed techniques exploit features extracted from individual management messages and from groups of messages called bursts. Moreover, other experiments show what happens when one or more features are split into their components or when the logarithm of their value is used. Before running the algorithm, a feature selection was performed and exploited to improve the accuracy. The clustering methods considered are DBSCAN and OPTICS.