New technology
AI and computer vision for workplace safety: capabilities, limits and governance
Cameras that can recognise hazards are spreading across sites. A sourced look at what the technology is, where it helps and how to manage AI risk.
Computer vision has been used in factories for decades to inspect parts. What is new is the use of AI models on ordinary site cameras to spot unsafe conditions: a person in an exclusion zone, a missing hard hat, a blocked fire exit, a forklift near pedestrians. Vendors make bold claims. This article separates the established technology from the hype and sets out a governance approach grounded in published frameworks.
Machine vision: the established foundation
Machine vision is the technology and methods used to provide imaging-based automatic inspection and analysis for such applications as automatic inspection, process control, and robot guidance, usually in industry.2 Its output is information rather than a picture: The information extracted can be a simple good-part/bad-part signal, or more a complex set of data such as the identity, position and orientation of each object in an image.2
The broader field of computer vision includes applications that matter to safety teams, such as Detecting events, e.g., for visual surveillance or people counting1 and Analyzing the condition of facilities in industry or construction.1 Industry is the largest application area: For 2024, the leading areas of computer vision were industry (market size US$5.22 billion), medicine (market size US$2.6 billion), military (market size US$996.2 million).1
Where vision AI can help safety
- Exclusion zones and vehicle interaction. Detecting people near moving plant, robots or forklifts and triggering warnings.
- Housekeeping and egress. Flagging blocked walkways, exits or fire equipment for follow-up.
- PPE observation. Estimating how often required PPE is worn in a defined area, as a leading indicator.
- Condition monitoring. Spotting leaks, smoke or damaged guarding on fixed equipment.
The common thread is that vision systems are good at repetitive observation over long periods — something humans do poorly. They are much weaker at judging context: whether a task was authorised, whether a deviation was justified.
Limits you should plan for
False positives and negatives. Every model misclassifies some scenes. Too many false alarms create fatigue; missed detections create false confidence. Measure both on your own footage before relying on alerts.
Control level. In the hierarchy of controls, an alert that asks a person to change behaviour is administrative. It does not replace guarding, segregation or interlocks. Use vision analytics to find where higher-level controls are needed.
Surveillance concerns. The same worry raised about wearables applies here: data can be repurposed. The encyclopaedic summary of wearables notes that employers could theorectically repurpose the data that wearables collect for objectives other than health such as worker surveillance.4 Camera analytics deserve the same scrutiny.
Governing AI with the NIST AI RMF
NIST’s AI Risk Management Framework offers a neutral structure. It is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems.3 It was Released on January 26, 20233, and NIST notes that The AI RMF 1.0 is being revised as part of the White House AI Action Plan.3 For industrial users, it is relevant that On April 7, 2026, NIST released a concept note for an AI RMF Profile on Trustworthy AI in Critical Infrastructure.3
Applied to a safety vision project, the four functions translate into practical steps:
- Govern: name an owner, agree the purpose, and set rules for data access and retention with worker representatives.
- Map: document the use case, the cameras, the people affected and what decisions the outputs will inform.
- Measure: test detection accuracy on your own scenes, lighting and PPE colours, and track false alarm rates over time.
- Manage: decide how alerts are handled, how errors are corrected, and when the system should be switched off or retrained.
Questions to ask vendors
We have not benchmarked vision vendors, so we suggest criteria instead: where processing happens (on site or cloud); whether faces are blurred or not stored; how accuracy was measured and on what data; how models are updated; integration with your incident and action systems; and whether the vendor maps its controls to a framework such as the NIST AI RMF.
A pilot plan you can reuse
Week 0–2: scope. Choose one area and one hazard, agree success measures, consult workers and publish a short notice explaining what the cameras analyse and what they do not.
Week 2–6: shadow mode. Run detection without alerts and compare results with human review of the same footage. Record true detections, false alarms and missed events.
Week 6–10: live alerts. Switch on alerts to a named responder, track response times and collect feedback from the crew.
Week 10–12: decide. Decide whether to scale, change the scope or stop, based on measured performance and the engineering controls the data revealed. Document the decision under the Govern function.
A pilot that ends with a new barrier or a redesigned traffic route has succeeded even if the camera system is later switched off.
Machine vision background
Machine vision is the technology and methods used to provide imaging-based automatic inspection and analysis for such applications as automatic inspection, process control, and robot guidance, usually in industry.2 Safety applications build on the same cameras and models but face harder problems: people move unpredictably, lighting changes, and errors affect individuals.
Risk management frameworks
NIST's AI Risk Management Framework is a common reference. It continues to evolve: on April 7, 2026, NIST released a concept note for an AI RMF Profile on Trustworthy AI in Critical Infrastructure.3 Use such frameworks to document intended use, risks, testing and monitoring.
Pilot evaluation metrics
- True and false alert rates, checked by humans on a sample.
- Response time from alert to action.
- Changes in observed unsafe conditions over time.
- Worker perception surveys.
- Data protection compliance review.
Key takeaways
- Machine vision is mature for inspection; safety analytics on site cameras is newer and must be validated locally.
- Use alerts to find where engineering controls are needed rather than as the control itself.
- Govern the project with a recognised framework and agree privacy rules with workers first.
Frequently asked questions
Is computer vision the same as machine vision?
Machine vision is the applied engineering discipline used mainly in industrial inspection and guidance. Computer vision is the wider scientific field. Safety analytics draws on both.
Can AI cameras prove PPE compliance?
They can estimate how often required PPE appears to be worn in a camera’s view, which is a useful leading indicator. Accuracy varies, so validate on your own footage and avoid using results for individual discipline without human review.
Is the NIST AI RMF mandatory?
No. NIST describes it as intended for voluntary use. It is widely used as a reference structure for AI governance.
Where should we start?
Pick one well-defined hazard, such as pedestrian and forklift interaction in a single area, run a time-limited pilot, and measure false alarms and missed events before scaling.
Does vision AI replace inspections?
No. It adds continuous observation, but competent people still need to inspect, judge context and verify controls.
Should alerts discipline workers?
Most experts advise against it. Use aggregated data to fix conditions and coach, and agree the rules with workers and their representatives beforehand, otherwise trust and reporting suffer.
Do we need consent?
It depends on local data protection and labour law; get legal advice before deployment.
Sources
Every factual statement marked with a number links to the source below. Sources were retrieved on 23 October 2026; check the publisher for later revisions.
- Computer vision (Wikipedia) en.wikipedia.org
- Machine vision (Wikipedia) en.wikipedia.org
- AI Risk Management Framework (NIST) nist.gov
- Wearable technology (Wikipedia) en.wikipedia.org
