Enhancing Crowd-Monitoring Through WiFi Fingerprint Analysis

The proliferation of smartphones, IoT devices, and other modern technologies has transformed cities into interconnected ecosystems, generating vast amounts of data. Accurately estimating crowds and counting people has become crucial for urban planners, transportation managers, and security agencies. By leveraging real-time data from various sources, decision-makers can optimize resource allocation, enhance security measures, improve customer experiences, and create more efficient urban environments. Capturing and analyzing network traffic has emerged as a valuable method for accurately estimating people's presence in specific areas. WiFi and Bluetooth are the two main types of signals that can be inspected, with WiFi being the preferred option for privacy reasons. WiFi Probe Requests, emitted by devices when they search for available WiFi networks, provide valuable data on the number and movement of people in specific areas. The main focus of this thesis is on estimating people counts through capturing, processing, and analyzing these kinds of messages. Firstly, a synthetic Probe Requests generator was developed to replicate Probe Requests sent by a customizable number of different devices. The generator is designed to simulate realistic Probe Request traces, based on data from real case scenarios. By leveraging this simulator, it is possible to provide a precise ground truth reference for the number of devices present in a given area. This approach enhances both the evaluation phase of counting methods and the training phase for machine learning techniques. Secondly, crowd-monitoring techniques have been employed to address the challenge of people counting. Since the probe requests' MAC address is randomized by modern operating systems, counting based solely on different addresses is not feasible. Instead, the focus has shifted to handling complex data patterns and extracting meaningful insights from messages. Based on an in-depth understanding of probe request fields and time features, two advanced frameworks have been successfully developed. The parameters of the algorithms have undergone rigorous training using vast amounts of data generated by the Probe Requests generator, ensuring optimal performance and accuracy. Drawing inspiration from clustering methodologies, our systems adopt a similar approach to analyze probe requests. By leveraging the power of these techniques, the frameworks can categorize and group probe requests based on their source device or vendor model. For one of the two implemented systems, the messages' time features have been taken into account to estimate the number of devices present. The developed methods provide a good approximation for estimating the crowd in a given area, demonstrating that these approaches closely align the retrieved number of people with the ground truth provided by the generator. The outcomes of this research have practical applications across various domains. In retail analytics, accurate people counting and flow estimation provide insights into customer behavior, optimizing store layouts, staffing, and marketing strategies. Another important aspect is to enhance safety in public environments such as events, pedestrian traffic in urban areas, and emergencies. Additionally, the development of a synthetic Probe Requests generator contributes to the advancement of crowd-estimation techniques, providing a valuable tool for evaluating and improving counting methods based on WiFi probe requests.