Privacy Preserving Edge AI in Real-Time Video Analytics for Monitoring Systems

This paper introduces an edge-centric, privacy-preserving video analytics pipeline that transmits compact pose and appearance metadata instead of raw frames to reduce latency, bandwidth, and privacy exposure while maintaining actionable spatial analytics. The on-node stack combines lightweight detectors and pose estimators (YOLOv8pose, HRNet), ByteTrack tracking, and low-dimensional re-identification embeddings (OSNet); a server-side module performs cross-camera association, behavioral scoring, bird’s-eye aggregation, and query-backed spatial analytics. This study evaluates the system on a 16-camera testbed using the CHAD dataset plus live streams, measuring per-batch latency, throughput, and Physical-Cyber-Physical (PCP) latency against a cloud baseline. Empirical results demonstrate clear edge benefits for small-to-moderate node counts (e.g., PCP latency ≈4.7 s at 4 nodes vs ≈21.4 s for the cloud baseline) and reveal supra-linear latency degradation as node count and crowd density increase (notably at ⩾ 12 - 16 nodes). Long-duration trials highlight index-growth-induced latency drift, which an automated archival/reset policy reduces by approximately 15-20% median latency. Ablation studies quantify tunable accuracy-compute trade-offs stemming from estimator and embedding choices. Overall, the proposed pipeline offers a practical, configurable balance between responsiveness and detection coverage for latency-sensitive monitoring and discusses operational deployment considerations and privacy-policy compliance. Future work will explore adaptive load balancing, model compression/pruning, hierarchical database sharding, and multi-venue validation.

Hali tarjima qilinmagan