PPE and site safety
Helmet, vest and glove compliance per person, with alerts the moment someone is missing gear. New PPE items are added at runtime on an open-vocabulary detector, no retraining.
What we build
Computer Vision Development services
Restricted-zone alerts
Draw a polygon on the live feed and get an alert, with a saved snapshot, when a person crosses into it. CPU-optimised YOLOv11, so it runs on an ordinary server.
Theft and loss prevention
Custom YOLO models fine-tuned on shop-floor footage that watch drawers and hands, and save a clip of every flagged event for review.
Face-recognition attendance
Touchless check-in and check-out from a webcam, IP camera or phone, with ArcFace embeddings, automatic in/out detection and daily and weekly reports.
Action and behaviour recognition
Classify what people are doing over time from pose keypoints: walking, running, falling. Useful for safety, elderly care and incident detection.
Object and cash monitoring
Recognise specific objects such as AED, USD and EUR banknotes, and raise an email alert when cash is left unattended past a threshold.
Why camera AI fails in production, and how we avoid it
Most computer vision demos work on clean benchmark images and fail on a real site: glare at 2 pm, dust, a camera mounted too high, a worker half behind a pillar. We design for the footage you actually have. Before writing any model code we pull a sample of your recordings, label what matters, and measure how a baseline detector performs on it. That number decides the approach, not a slide.
Where the list of things to detect will change, such as PPE items or products on a shelf, we use open-vocabulary detectors like YOLO-World and OWL-ViT, so your team can add a class by typing it rather than waiting for a retrained model. Where accuracy and speed on a fixed set matter more, we fine-tune a YOLO model on your own footage. We explain the trade-off in our post on YOLO versus open-vocabulary detection.
How a project runs
- Footage review, week 1. We review sample clips from your cameras, agree the events that matter and the alert you want, and measure a baseline.
- Pilot, weeks 2 to 3. One camera, one use case, live alerts on your own feed. You see real true and false positives before committing further.
- Rollout. More cameras and use cases, an operator console, alert routing to email, WhatsApp or your existing system, and retention rules for clips and images.
- Run and improve. We log missed and false alerts, retrain or re-prompt on them, and report accuracy monthly.
Built for UAE sites
Our safety work is written with Dubai conditions in mind: heat haze, dust, night lighting and high-mounted cameras. We deploy on-premise or on UAE-hosted servers when footage cannot leave the site, blur faces where identity is not needed, and keep images only as long as your policy allows. Read how this works on a live site in computer vision for construction-site safety in the UAE.
Pricing and timelines
A fixed-price pilot on one camera and one use case, two to three weeks, typically AED 12,000 to 18,000. Larger rollouts are scoped after the pilot. Every engagement starts with a written scope, a fixed first step and a delivery date.
Technology we use
Questions
Do we need new cameras?
Usually not. Our systems read standard RTSP or MJPEG streams from existing IP cameras and NVRs. We only recommend new hardware when a camera's angle or resolution makes the use case impossible.
Do we need a GPU server?
Not for most single-site deployments. Restricted-zone and PPE monitoring run on CPU-optimised models. Many cameras at high frame rates, or several heavy models at once, may need a GPU, and we size that in the pilot.
How accurate is it?
We measure precision and recall on your own footage during the pilot and report the numbers before rollout. Accuracy depends on camera placement and lighting, so we do not quote a figure before seeing your feed.
Where does the footage go?
It stays on your premises or on a UAE-hosted server you control. Alert images and clips are kept for the period you set, and faces can be blurred when identity is not required.
How much does it cost?
A two to three week pilot is typically AED 12,000 to 18,000 at a fixed price. Rollout pricing depends on the number of cameras and use cases and is agreed after the pilot.
Other services: Machine Learning & MLOps · Generative AI & LLM Applications · AI Visibility (GEO)
Planning a computer vision development project?
Tell us the problem and the data or cameras you have. We reply within one business day with a first step, a price and a date.