Computer Vision for Industrial Safety: From Cameras to Measurable Risk Reduction
Бекзат Маратұлы · June 6, 2026 · 8 мин
Answer capsule. HSE video analytics means neural network models (YOLO-class detection and tracking) running on top of the enterprise's existing cameras and flagging violations in real time: missing PPE, entry into a danger zone, hot work violations. Video is processed on edge servers inside the perimeter — footage never goes to the cloud. The key project metrics are detection recall and precision per violation type, and a false alarm rate low enough that dispatchers do not switch the system off. A pilot on 5–10 cameras takes 8–12 weeks and is calibrated to the specific site's conditions: dust, frost, night shifts.
Most industrial sites already have cameras — but they are used as an archive for post-incident review. Computer vision turns the same video streams into a prevention tool: the system sees a violation as it occurs and warns before it becomes an incident.
What violations video analytics detects
- PPE: hard hat, safety glasses, gloves, high-visibility vest, fall-arrest harness at height — tied to the zone (workshop requirements differ from outdoor areas).
- Danger zones: personnel entering a crane operating area, walking under suspended loads, entering vehicle operating radii; barrier and fencing control.
- Hot and gas-hazardous work: presence of a fire watch, firefighting equipment, compliance with the work permit.
- Vehicles and machinery: on-site speed, dangerous proximity between machinery and people, blind-spot monitoring for mining trucks.
- Worker condition: a person has fallen and is not moving (man-down) — critical for remote, low-staffed facilities.
Architecture: why video never leaves the site
Production site video is sensitive data: it shows people, processes and the facility perimeter. Inference therefore runs on GPU edge servers inside the enterprise network (Jetson-class devices or rack GPU servers). Only events reach the control room: the violation frame, timestamp, camera, type. This simultaneously resolves the security review and reduces bandwidth requirements — relevant for remote fields.
The metrics a system should be accepted on
| Metric | What it shows | Why it matters |
|---|---|---|
| Recall | The share of real violations the system caught | A missed violation is a failed safety function |
| Precision | The share of alerts that were real violations | Under a flood of false alarms, dispatchers stop reacting |
| False alarms per shift | The operational load on the dispatcher | The practical threshold for system adoption |
| Event-to-alert latency | Reaction speed | Prevention is a matter of seconds |
An important note on honesty: "99% accuracy" from marketing materials means nothing without the measurement conditions. Metrics must be measured on the specific site's video — with its lighting, dust and winter fog — and recorded in the pilot report.
How implementation works
- Audit (2–3 weeks). Camera and zone survey, selection of the 3–5 highest-risk violation types, alignment with the HSE and security teams.
- Pilot (8–12 weeks). Fine-tuning models on site footage, deploying edge inference on 5–10 cameras, calibrating thresholds together with dispatchers.
- Evaluation. A report with metrics per violation type, violation dynamics over the period and an effect calculation.
- Scaling. Extension to remaining cameras and zones, integration with permit-to-work systems and HSE reporting.
105 Industrial AI ("105kz" LLP, an Astana Hub resident) builds industrial safety video analytics for Kazakhstan's enterprises — on existing cameras, inside a closed perimeter, with interfaces in Russian and Kazakh. We start with an audit and a pilot under NDA. Contact us — we will show how we measure metrics and effect.