A camera system shares edge work and adapts models centrally

Can surveillance systems process many continuous camera streams on limited local computers while accounting for how scenes differ and change?

A model trained on fixed data may not work reliably across every camera, so the system must manage local computing demands and scene-specific adaptation together.

YUTELA shares edge processing across camera streams and retrains a separate model for each camera on a central server.

The paper asks whether surveillance cameras at different sites can share limited local computing while adapting to the distinct conditions in each scene.

YUTELA addresses this with shared edge processing and separate camera models that are retrained centrally from recent samples identified as normal.

In controlled video tests, adding TensorRT improved YUTELA’s processing rate and latency on the same edge device. Optimized YUTELA also processed more frames per second than the workstation-based Ancilia system in seven of nine settings, although that comparison changes both hardware and software and latency remained higher in some conditions. A 170-hour deployment showed the complete system operating across four sites.

The deployment measured system operation and alert timing, but not whether retraining made anomaly detection more accurate. The alert-time range also reflects different site conditions, so it is not a direct comparison between sites.