SafetyPro
  1. Home
  2. Blog
  3. AI Safety Scanner: hazard and PPE detection from site photos and CCTV
Blog

AI Safety Scanner: hazard and PPE detection from site photos and CCTV

A supervisor takes one photo of a work area and, in about a minute, gets a list of unsafe acts, unsafe conditions and PPE gaps, each marked on the image with a severity and a suggested fix. The same analysis can run on CCTV frames. This article explains exactly what the SafetyPro AI Safety Scanner does, what it returns, and where its limits are.

SafetyPro Team · · 8 min read

The problem: hazards in plain sight

Most serious hazards are not hidden. A missing guardrail, a worker without a harness, a blocked fire exit or a trailing cable is visible to anyone who looks. The problem is that people stop seeing what they walk past every day, and inspectors cannot be on every work front at once.

Photos are already everywhere on site, but they usually end up in chat groups and phone galleries. Nobody reviews them in a structured way, and nothing links what the photo shows to an owner and a deadline. What good looks like is simple: every photo or camera frame can be checked against the same checklist, findings are recorded with evidence, and the serious ones become tracked actions the same day.

How a photo scan works

The scanner is a module of the SafetyPro web platform, so it runs in any browser, including a phone or tablet browser on site. The flow has four steps:

  1. Take a photo or upload one. JPG, PNG and WebP are accepted, up to 10 MB per image.
  2. Optionally choose the work zone: construction site, workshop or manufacturing, warehouse or logistics, working at height, confined space, chemical storage, welding or hot work, or office. You can also leave it on auto-detect and add a location name.
  3. Press Analyze Scene. The image is sent to a vision AI model with a structured HSE inspection brief.
  4. Review the result on the same screen: risk score, findings table, marked-up image and recommendations.

The zone matters. When you pick one, the AI is told which PPE that zone requires. For a construction site that is hard hat, safety glasses, hi-vis vest, steel-toe boots and gloves, plus checks on scaffolding, fall protection and excavations. For confined space entry it looks for a gas detector, harness, rescue equipment, an attendant and a visible entry permit. For welding it looks for a welding shield, fire-resistant clothing, welding gloves, a nearby extinguisher and a visible hot work permit.

What the AI returns

Each analysis produces a structured result, not just a paragraph of text. It contains:

  • An overall risk score from 0 to 100 and a risk level: safe, low, medium, high or critical.
  • The number of people detected and a PPE compliance percentage for the scene.
  • A per-person PPE check covering hard hat, safety glasses, hi-vis, gloves, safety boots and harness. People are numbered left to right in the image, so a finding can point to Person 2 rather than to a name.
  • A list of findings. Each one has a category, a description, a severity (critical, high, medium or low), where it is in the image, a reference to the relevant OSHA standard, ISO clause or good practice, and a specific recommended corrective action.
  • A two to three sentence summary and the top three priority recommendations for the area.

Findings are sorted into 10 categories: unsafe act, unsafe condition, PPE violation, ergonomic risk, housekeeping, fire safety, electrical, chemical, working at height and confined space. For ergonomic findings the AI is asked to reference REBA, RULA or NIOSH methods where they apply.

The most visual part is the annotated scene. Every finding and every person is drawn as a box on the photo, colour coded by severity, so a supervisor can see at a glance where the problem is. Markers can be toggled on and off, and the marked-up image can be downloaded for a toolbox talk or a report.

From finding to action in the Action Tracker

A finding only matters if someone fixes it. Under the findings table, Create Actions from Findings opens a review screen with one card per finding, all selected by default. For each card the platform pre-fills:

  • The action title, taken from the recommended corrective action and editable.
  • The priority, mapped from the AI severity: critical, high, medium or low.
  • A suggested due date: 1 day for critical findings, 3 days for high and 7 days for the rest.
  • A responsible person, chosen from your active users.

You untick what you do not agree with, adjust the rest, and press Create Selected Actions. Each action lands in the Action Tracker as an open corrective action with the source set to AI Safety Scanner, a description that records the finding, category, severity, regulation reference and recommendation, and a link back to the analysis that produced it. By default the scan photo is attached to every action as evidence. Nothing is created until a person confirms it.

History, trends and reports

Every completed scan is saved against the current project. The History tab lists scans by date, source, risk level, score, number of findings and PPE compliance, filtered to a single day, the last 7 days or all time. Any scan can be reopened with its full findings and image.

The Trends tab turns scans into management information: total scans, total findings, critical scans and active cameras, a risk distribution, the top finding categories, an analysis timeline and a risk score trend over time. This shows whether housekeeping or PPE findings are actually going down after a campaign.

Export PDF produces a scene analysis report with the risk score, people and PPE figures, the annotated image, the AI summary, the priority recommendations and every finding with its regulation reference and corrective action. If your organisation has published a report template for the scanner in the Report Template Engine, the PDF uses your branded cover page.

Camera and video analysis

The same analysis engine can work on frames from site cameras. Cameras are registered in the CCTV Cameras tab with a name, location, zone type and connection details, and camera passwords are masked in the stored stream address. Only users with create, edit or delete rights for the scanner can add, change or remove cameras.

For cameras connected through Imou Cloud, two ways of analysis are available today:

  • Scheduled background scans. You set active days, a start and end time and a capture interval. A scheduler on SafetyPro servers checks every 2 minutes and scans each camera that is inside its window and due, even when nobody has a browser open.
  • Live View, with a snapshot every 3 seconds or HLS live video where the camera supports it. From Live View you can analyse the current frame on demand, or switch on auto-analyze at an interval between 15 and 600 seconds while the view is open.

Every camera scan uses the camera's zone type for PPE context, appears in History and Trends with the other scans, and can be turned into tracked actions in the same way as a photo. If your cameras use a different platform or a local network recorder, talk to the SafetyPro team about the right way to connect them.

Live ergonomics in the browser

The scanner also has a Live Ergonomics tab. It opens the device camera and runs pose estimation directly in the browser, tracking 33 body landmarks. From the joint angles of the neck, trunk, upper arm, lower arm and knee it calculates a REBA style score in real time, from negligible (1) to very high (11 and above), with the matching action level. A snapshot with the pose overlay can be saved to the device. Because the pose model runs in the browser, the live video for this tab is not sent to SafetyPro.

Treat this as a quick screening tool for manual handling and awkward postures. It uses a simplified REBA calculation, so a formal ergonomic assessment is still the right step for high scores.

Safeguards, privacy and limits

The scanner is designed to help people see more. It does not make decisions on its own. Some points worth knowing:

  • AI output is a suggestion. The model is told to flag anything potentially unsafe and to say unable to confirm when the image is unclear, so expect some cautious findings that a competent person should review.
  • Actions are never created automatically from a photo scan. A user selects, edits and confirms them.
  • People are counted and numbered in the image. The scanner does not try to identify who they are.
  • Each organisation's scans, findings and cameras are kept separate. Photos are stored in your organisation's own storage area and shown through time-limited signed links.
  • An analysis has a 55 second time limit. If the AI service is busy, the platform retries and then shows a clear try again message instead of a raw error.

A practical rollout plan

  1. Start with photos. Ask supervisors to scan one photo per work front per shift for two weeks, always selecting the correct zone type.
  2. Agree a review rule: critical and high findings are reviewed the same day and turned into actions; medium and low findings are discussed at the toolbox talk.
  3. Use the Trends tab weekly to pick one category to target, such as housekeeping or PPE, and track whether it falls.
  4. Add cameras where the risk is highest and supervision is thinnest, such as loading bays, scaffolds and hot work areas, with active hours that match the shift pattern.
  5. Review critical camera findings every day with the supervisors of the areas they cover.

HSE insights and practical guides

See it on your own projects

Tell us about your sites and we will show you how SafetyPro fits your HSE operations.