Real-Time Video Surveillance Analysis at the Edge in Smart Cities

The number of video surveillance cameras is also increasing as the number of smart cities is soaring, driven by the need to improve residents' safety. Traffic is designed and cities are planned based on statistics. Some issues in traditional cloud-based surveillance systems include high latency, bandwidth congestion, privacy breaches, and scalability limitations. To defeat them, in this paper we present the Federated Edge-Cloud Collaborative Video Surveillance Framework which is based on three core innovations: (1) Edge AI to offer on-device video analytics in real-time, (2) Federated Learning, which offers privacy-preservation and distributed training of the model without transfer of raw video, and (3) Cross-domain event correlation, which correlates video with other sensor data to offer more context. The framework achieved the total reductions of up to 92 percent of latency (latency 25 ms under 100 edge devices in a simulated innovative city situation), 80 percent bandwidth, and 93.5 percent mAP (mean Average Precision) of mAP (mean Average Precision) that is comparable to cloud model-based performance with 100 edge devices. Besides, intelligent resource distribution and task unloading resulted in a 40% reduction in energy consumption (24 Wh vs. 30 Wh for edge-only equal-load loads). It is also highly GDPR-compliant, as the system does not require transferring raw video data to other parties; only encrypted metadata or model updates are transferred. Generally, the framework is a scalable, power-saving, privacy-friendly solution that can be adjusted to the changing objectives of intelligent city monitoring.

Hali tarjima qilinmagan