On-prem machine-learning inference inside the video wall pipeline: anomaly detection, automatic source promotion, face / plate counting — without sending feeds to the cloud.
What it is
Edge AI in this context means running ML models directly on the video wall controller, against the live source streams, without round-tripping pixels through a cloud service. The platform consumes framework-agnostic models (typically ONNX) and exposes inference results to the layout engine so the wall can react to events in real time.
What you actually run
- YOLO v8 / v9 via ONNX for object, person, and vehicle detection — well-supported on commodity NVIDIA and Intel GPUs through OpenVINO or TensorRT.
- Anomaly-style models — compares the current stream against a historical baseline, raises a tile when the deviation crosses a threshold. Cheap, surprisingly effective for unattended monitoring rooms.
- Source promotion — the most operator-visible feature: when a detection fires, the relevant tile grows, moves to the centre of the wall, and stays there until acknowledged.
Why on-prem matters
For NOC / SOC, situation rooms and any regulated facility, sending live video feeds to a cloud inference endpoint is a non-starter: latency, bandwidth, sovereignty, and (in the EU and Russia) statutory constraints all push the model down to the edge. The 2025-2026 product moves from Userful (Infinity EdgeAI with Microsoft and NVIDIA) and Visiology (Cortex, RU) confirm this as the category direction.
What it costs
Single-GPU workstations (NVIDIA RTX A4000 / A5000 class) absorb YOLO-class detection across 8-12 simultaneous 1080p streams without choking the compositor. Beyond ~20 streams the AI workload starts competing with the rendering workload, and the right architecture is a separate inference node beside the wall controller.
Common pitfalls
- Treating AI as a feature toggle. A model that drives layout has to be tuned per site and audited — false positives promote noise to the centre of the operator's attention and undermine trust in the wall.
- Confusing inference (the model running on live frames) with training (which happens elsewhere on a curated dataset). Production walls run inference only.