The monitoring of large crowds is essential to optimize traffic flows, ensure safety at large-scale events, and plan effective evacuation routes during emergencies. However, such monitoring rightfully leads to privacy concerns, especially when tracking individuals rather than groups. Existing approaches attempt to address these concerns by pseudonymizing personally identifiable information and restricting the analysis to statistical counts. However, these methods fail to preserve privacy, particularly when small groups can be correlated with external data. To combat this issue, we leverage the idea that crowd monitoring applications are interested in only large crowds (e.g., > 100 people) and can deal with low noise levels (e.g., it does not matter whether we count 95 or 105 people). We propose and evaluate two methods that not only protect individual data, but also enhance privacy by introducing varying levels of controlled noise: higher for smaller groups and lower for larger crowd movements. These methods include probabilistically: (1) sampling hash functions and (2) sampling detected identifiers. We show that our methods significantly reduce the risk of re-identification in small crowds while maintaining high precision in large crowd estimations, making them highly effective for privacy-preserving crowd monitoring.
Author: Fatemeh Marzani
Publication date: 2025