What our AI actually sees on your cameras.
Virtual Sites Security runs a YOLO-family object-detection model on every monitored camera stream. These frames are real model output on night scenes — people and vehicles boxed with confidence scores, weather and lighting ignored, and one event escalated to a live operator.
Detection, zones, then a human.
The model does the looking; our operators do the deciding.
1 · Classify every frame
The detector scans each camera several times a second and labels people, cars, trucks and motorcycles with a confidence score. Trees, rain, animals and headlight glare produce no label — so they produce no alert.
2 · Apply your zones & schedule
A detection only matters where and when you said it does: inside the fence after 8 p.m., at the dock after the night shift, never on the public sidewalk. Zones and arming schedules are drawn with you during setup.
3 · Operator verification
Every zone hit opens the live stream on an operator desk in Los Angeles. The operator confirms what the subject is doing, speaks through the on-site speaker and dispatches response — the model never acts alone.
Real detections on eight night scenes.
Green and blue boxes are classified objects with the model's confidence score. The red box is the detection that crossed an armed zone and was sent to an operator.
The detection pipeline, in plain terms.
Everything below is configured per camera during the two-night shadow-mode test, so thresholds match the lighting and traffic at your site.
| Model family | YOLOv8 (medium), fine-tuned for low-light scenes |
| Classes escalated | Person, car, truck, bus, motorcycle |
| Classes ignored | Animals, vegetation, weather, shadows, headlights |
| Confidence threshold | 0.45 default; tuned per camera |
| Zone rules | Intrusion, line-cross (directional), loiter / dwell, crouch, group |
| Inference location | Edge appliance on site or VSS cloud — no change to your cameras |
| Thermal support | Yes — same pipeline on thermal streams |
| Operator review | Every zone event, on live video, within 30 seconds |
About the AI
No. The model filters thousands of motion events down to real people and vehicles in your zones; a licensed operator reviews every one of those on live video before anything happens.
Yes for most IP cameras. The detector runs on the video stream, not inside the camera, so no camera firmware or vendor features are required.
No. The detector classifies objects (person, vehicle) — it does not perform facial recognition. Plate reading is a separate, opt-in LPR service with its own policy.
Night is what the model is tuned for — all of the samples on this page are night scenes. Heavy fog or rain lowers range, which is why we set per-camera thresholds and tell you where IR or thermal cameras are needed.
Get a flat monthly price for your property.
Send a site address and a camera count. We come back with a coverage map, a recommended program and pricing — usually within one business day.
Tell us about your site
Sent to info@virtualsitesecurity.com · We reply within one business day. No obligation.
See the model run on your own cameras.
We can connect to an existing camera and show you live detections during the first call.