GeoPrivd: Geospatial Privacy-Preserving Urban Traffic Video Analytics Queries With Trajectory-Level Sensitivity Control

Abstract

Urban-scale video analytics systems hold significant promise for traffic monitoring, pedestrian flow estimation and public safety. However, the acquisition and processing of spatio-temporal data extracted from traffic videos pose significant challenges to individual privacy. Existing approaches often rely on trusted tracking pipelines or expose fine-grained data, making them incompatible with strong privacy preservation. In this paper, we introduce GeoPrivd, a geospatial urban traffic video analytics framework that enables declarative spatio-temporal queries over object observations with formal differential privacy guarantees. GeoPrivd enforces privacy without relying on raw trajectories or trusted trackers, using a trajectory slicing and bounding mechanism to limit per-user influence. Queries are expressed in GeoPrivdQL, a high-level domain-specific language with built-in privacy controls. Through static query validation and calibrated noise injection, GeoPrivd ensures privacy-preserving analytics results across diverse analytics tasks. Experiments on real-world urban traffic datasets demonstrate that GeoPrivd consistently outperforms standard differential privacy (DP) baselines in utility, stability, and privacy resilience.