Skip to main content

On-edge processing

Smart cameras and AI boxes move part of the recognition pipeline onto the device itself. Instead of shipping raw video to the server, the device runs face detection — and optionally face-recognition tasks such as template extraction — on its own chip, then sends only the results. The server still coordinates matching and events, but it no longer carries the full video-processing load for that camera.

This shifts the economics of a site. Because only metadata and selected frames travel the network, edge devices use dramatically less bandwidth than full RTSP streams, and the server needs far less CPU/GPU per camera. The result is lower latency and lower server-side hardware cost, which compounds as a deployment grows from a handful of lanes to dozens. This is why on-edge processing keeps gaining ground even though RTSP remains fully supported.

Edge processing depends on the device, so the hardware list is narrower and more specific than for RTSP. VPP's embedded stream processor runs on smart cameras built around the Ambarella SoC, and on AI boxes using the Hailo-8 accelerator or NVidia Jetson modules. The exact tested models are maintained in the platform documentation and change over time, so always confirm against the current list before purchasing — see supported devices below.

Choose this approach when you are scaling to many cameras, your network or server budget is constrained, or you need the lowest possible latency. The trade-off is less vendor freedom and a dependency on the device's processing capability, so verify a model is on the supported list before committing.

Compatible hardware

The authoritative, regularly updated list of tested smart cameras and AI boxes lives in the VPP embedded documentation. Devices outside the list may still be compatible — if in doubt, confirm with Innovatrics before procurement.