Detecting human encounters from WiFi radio signals

Geert Vanderhulst, Afra J. Mashhadi, Marzieh Dashti, Fahim KawsarMum 2015 · DOI

TL;DR

This paper introduces a novel framework for automatically detecting spontaneous human encounters using ubiquitous WiFi signals from mobile devices. By analyzing the spatio-temporal properties of WiFi probe requests and applying a sociologically-grounded model, the system achieves high accuracy in identifying short-lived, informal interactions without requiring dedicated hardware or significant user effort.

What Problem Does This Paper Address?

Understanding human social interactions is crucial for fields ranging from social psychology to urban planning. While digital traces from social media and communication logs offer insights into online behavior, capturing real-world, face-to-face interactions has been challenging. Existing methods often rely on expensive, intrusive technologies like video surveillance or wearable sensors, which raise privacy concerns and limit scalability. The paper aims to bridge this gap by leveraging the pervasive nature of WiFi-enabled mobile devices to passively and unobtrusively detect "spontaneous encounters" – short, informal interactions between individuals in physical spaces.

Methods

The core of the framework relies on a network-centric approach using existing WiFi infrastructure.

  • Zero Mobile Sensing: The system avoids requiring users to install special apps or carry dedicated sensors. Instead, it leverages the "Probe Request" WiFi management frames that mobile devices periodically broadcast to discover networks.
  • Radio Signal Capturing: WiFi Access Points (APs) are configured to monitor for these "DIRECT-" probes, which are intentionally broadcast by a small companion app on the user's device. This allows for controlled data collection.
  • Copresence Detection: The system compares the received signal strength (RSSI) values from different devices captured by multiple APs. Devices with similar RSSI "fingerprints" within a defined similarity threshold (approximating 3 meters) are considered "copresent." This copresence data is aggregated over a 1-minute window to ensure sufficient detection.
  • Human Encounter Inference: A sociologically-inspired model is applied to the copresence data. This model considers two key variables: the duration of copresence and the size of the group. Encounters are defined as spontaneous interactions lasting at least 1 minute and involving at least two people. A "stability feature" is calculated to distinguish genuine encounters from mere temporary co-location (e.g., people in a queue). An exponential moving average is used to detect the formation and decomposition of encounters.
  • Adaptive Window: The duration of the sliding window used for feature calculation is dynamically adjusted based on the number of copresent individuals, informed by empirical observations of human interaction group dynamics.

Datasets

  • Controlled Experiments: The authors scripted various use cases (e.g., queues, workspaces) involving 2-6 people for different durations to simulate spontaneous encounters.
  • In-the-Wild Deployment: A 7-hour live deployment was conducted with 8 employees in an open office space. Participants also maintained a diary of their encounters to serve as ground truth.

Key Insights and Ideas

  • Ubiquitous WiFi as a Social Sensing Tool: The paper demonstrates that readily available WiFi infrastructure and mobile device signals can be repurposed for sophisticated social sensing, offering a low-cost and scalable alternative to dedicated sensors.
  • Sociology-Informed Modeling: Integrating principles from sociology regarding group dynamics and interaction rituals significantly improves the accuracy of detecting spontaneous encounters, moving beyond simple proximity detection.
  • The "Chattiness" of WiFi is a Feature, Not a Bug: The periodic broadcasting of WiFi probe requests, often seen as an energy drain or network noise, is ingeniously leveraged as a signal for human presence and interaction.
  • Balancing Granularity and Robustness: The use of a 1-minute aggregation window and an adaptive window size for encounter detection strikes a balance between capturing short-lived interactions and mitigating the effects of signal fluctuations and probe losses.

Key Results

  • Controlled Evaluation: The framework achieved an F-Score accuracy of 0.96 (precision=1, recall=0.92) using an adaptive window size. Decomposition delay was less than 2 minutes on average.
  • In-the-Wild Evaluation: The system detected 10 out of 17 self-reported encounters (58% recall) and identified 5 additional false encounters, resulting in an overall precision of 86%. The detected encounter durations were within a 10-minute offset of the ground truth.

Limitations and Future Work

  • Device Heterogeneity: While the system aims to mitigate signal fluctuations, differences in device hardware (e.g., WiFi chip, antenna) can still introduce inaccuracies.
  • Desk-Sharing Scenario: The framework struggled to differentiate between continuous copresence in a shared desk area and actual spontaneous encounters, leading to false positives in some controlled scenarios. The authors suggest a filter to detect stable RSSI fluctuations indicative of shared spaces.
  • Limited Participant and Device Variety: The evaluation was conducted with a small number of participants and a specific device model (Moto G phones). Future work should assess performance with a wider range of devices and in denser deployments.
  • Companion App Requirement: While minimal, the framework currently requires a small application on the user's device to initiate synchronized probes. Future work envisions this functionality being integrated into mobile operating systems.
  • Congestion in Dense Deployments: The authors acknowledge that in very dense environments, probe congestion could become an issue, suggesting the need for alternative robust techniques.

Summary updated

BibTeX
@inproceedings{vanderhulst2015detecting,
  author = {Geert Vanderhulst and Afra J. Mashhadi and Marzieh Dashti and Fahim Kawsar},
  title = {{Detecting human encounters from WiFi radio signals}},
  booktitle = {Mum},
  year = {2015},
  doi = {10.1145/2836041.2836050},
}
Figure