AI Worker Safety Vision Systems for Factories

AI Worker Safety Monitoring Manufacturing: What Factory Managers Need to Know

AI worker safety monitoring manufacturing systems represent one of the most consequential applications of computer vision in industrial environments. Every year, workplace accidents on factory floors result in serious injuries, lost working days, regulatory investigations, and — in the worst cases — fatalities that could have been prevented. The case for using technology to reduce these incidents is straightforward. The implementation, however, is anything but simple, because deploying cameras to monitor workers on a factory floor touches on some of the most sensitive legal, ethical and organisational territory in industrial AI.

This guide is written for health and safety managers, operations directors and HR leaders at Dutch manufacturers who are evaluating whether computer vision safety systems could help them reduce workplace incidents. It covers how the technology works, what it can and cannot do reliably, where the genuine boundaries of GDPR and employee privacy law lie, and what responsible deployment looks like in practice. It is general information, not legal advice — and we say that not as a disclaimer to be skipped, but because the specific legal position of any camera monitoring system in your workplace must be assessed with qualified legal counsel and, under Dutch law, with your works council (ondernemingsraad).

Crux Digits builds computer vision safety systems for Dutch manufacturers — PPE detection, hazard zone monitoring, ergonomics analysis — with privacy by design as a core architectural principle rather than an afterthought. We work across manufacturing environments including metal fabrication, chemical processing, food production, and logistics operations. Our approach starts with the safety outcome, works backwards to what data is necessary to achieve it, and builds in the minimum data footprint required.

How Does Computer Vision Improve Worker Safety on Factory Floors?

Computer vision safety systems on factory floors work by deploying cameras at defined positions — typically above hazard areas, at entry points to restricted zones, and along production lines where specific safety rules apply — and running a real-time AI model that analyses the camera feed to detect safety-relevant events. The key distinction from conventional CCTV is that the AI model does not record or review footage for general surveillance purposes: it is trained to detect specific safety conditions, and its output is an alert or signal, not a video archive of worker activity.

The main use cases that have demonstrated practical value in manufacturing environments fall into three categories.

PPE detection. Computer vision PPE detection systems identify whether workers in a defined area are wearing the required personal protective equipment — hard hats, high-visibility vests, safety glasses, gloves, steel-toed footwear, ear protection, respiratory masks. The camera analyses the worker's image, the model classifies the presence or absence of each required PPE item, and if a required item is missing, an alert is generated — a warning light, an audible alarm, or a notification to a supervisor. Critically, the system is detecting equipment compliance, not monitoring individual worker behaviour or productivity. That distinction matters enormously for both privacy law and employee relations.

Hazard zone and proximity detection. Vision AI hazard detection factory systems monitor access to defined danger zones — areas around heavy machinery, forklifts, chemical handling stations, or high-voltage equipment — and generate alerts when a person enters a zone without authorisation or when a person and a moving vehicle come within a defined proximity threshold. These systems replace or supplement physical barriers and audio warnings with an additional layer of real-time spatial awareness, particularly valuable in dynamic environments where hazard zone boundaries change with production configuration.

Ergonomics and posture monitoring. AI ergonomics monitoring production systems analyse body posture and movement patterns to identify ergonomic risk factors — awkward bending angles, repetitive overhead reach, sustained static postures — that are associated with musculoskeletal disorders over time. These systems are typically used for workstation analysis rather than continuous monitoring: a camera captures footage of a task being performed, the AI model analyses the posture data, and the output is an ergonomic risk report that informs workstation redesign. Continuous real-time ergonomic monitoring of individual workers raises significant privacy concerns (discussed below) and is not a model Crux Digits recommends for most manufacturing environments.

What Computer Vision Safety Systems Can and Cannot Do

Honest assessment of technology capabilities is essential before any procurement decision, particularly for safety-critical applications where overconfidence in an AI system can itself create risk.

What these systems can do reliably: Detect the presence or absence of high-visibility PPE items (hard hats, hi-vis vests, safety glasses) in good lighting conditions with well-trained models. Detect when a person crosses a defined boundary into a hazard zone. Detect proximity between people and moving objects when the camera placement provides adequate line of sight. Generate real-time alerts faster than human monitors could respond to live camera feeds. Provide aggregate data on the frequency and location of safety events — information that supports root-cause analysis and targeted safety improvement work.

What these systems cannot do reliably: Detect PPE that is worn incorrectly but present in the image — a hard hat worn backwards, safety glasses pushed onto the forehead, a respirator worn below the nose. Detect safety conditions that are not visible to the camera — a worn safety harness that appears intact but is past its inspection date, a tool guard that has been removed. Perform reliably in poor lighting, in heavy dust or steam environments, or where workers are frequently occluded by machinery or other workers. Substitute for a safety management system, a safety culture, or the expertise of a qualified health and safety professional.

These limitations are not arguments against the technology — they are constraints that should shape deployment design. A PPE detection system should be deployed at entry points and open areas where PPE compliance can be reliably observed, not in areas where it will generate excessive false alerts due to occlusion or poor visibility. The aggregate safety event data it generates should feed a broader safety management process, not replace one.

GDPR, Employee Privacy and the Legal Framework for Factory Camera Monitoring

This is the section that determines whether a computer vision safety deployment is lawful, and it deserves careful attention. Camera monitoring of workers on a factory floor involves the processing of personal data under the General Data Protection Regulation (GDPR), and the Autoriteit Persoonsgegevens (AP) — the Dutch data protection authority — has published clear guidance on what is and is not permissible. The Autoriteit Persoonsgegevens website is the authoritative source for Dutch employers on the legality of workplace monitoring.

The key principles that apply to any AI workplace safety compliance manufacturing camera system are as follows.

Proportionality and necessity. The monitoring must be limited to what is strictly necessary to achieve the stated safety objective. Cameras positioned to monitor a specific hazard zone or PPE compliance at a defined entry point are more likely to meet the proportionality test than cameras providing comprehensive coverage of all worker activity across the entire factory floor. Data retention periods must be as short as possible — for real-time alert systems, that may mean no permanent video storage at all. For incident investigation, retention may be justified for a limited period after an incident, not as a general archive.

Legitimate processing basis. Safety monitoring in a manufacturing environment may be based on legitimate interest (the employer's interest in preventing workplace injuries) or on legal obligation (compliance with the Working Conditions Act, the Arbowet). The chosen basis must be documented and the legitimate interest assessment must demonstrate that the privacy interests of workers do not override the safety objective. This is a genuine balancing exercise, not a formality — and the design of the system (what data is collected, how long it is retained, who has access, what it is used for) is what makes the balance credible or not.

Purpose limitation and function creep. A camera system deployed for safety monitoring must not be used for performance monitoring, productivity tracking, attendance management, or disciplinary proceedings unrelated to safety. This is not just a legal requirement — it is the most common way that well-intentioned safety monitoring programmes lose employee trust and turn into contested employment disputes. The system design should make function creep technically impossible where feasible: anonymising or blurring individuals in the general video feed while only retaining non-identifiable safety event data, for example.

Transparency and worker notification. Workers must be informed that camera monitoring is in place, what it is monitoring, why, and how the data is used. This is not optional and cannot be waived. Effective communication of the monitoring purpose — safety, not surveillance — also has practical importance for the programme's legitimacy in the eyes of your workforce.

The Role of the Works Council (Ondernemingsraad)

Under Dutch law, the Wet op de ondernemingsraden (WOR) gives the works council (ondernemingsraad, or OR) a right of consent (instemmingsrecht) over the introduction or significant modification of systems for monitoring or evaluating employees. A computer vision system that captures images of workers and generates alerts about their behaviour clearly falls within the scope of this right.

This means that before a computer vision safety incident prevention system can be deployed, the OR must be consulted and must give its consent. Attempting to deploy without OR consent is not just a legal risk — it is a practical one. A system that workers regard as having been imposed without consultation is far more likely to generate resistance, distrust and unintended consequences (such as workers avoiding monitored areas or wearing PPE only in camera sight lines) than one that was developed with genuine OR engagement.

In practice, productive OR engagement on a computer vision safety system involves:

• Presenting the specific safety problem the system is designed to solve, with incident data to support the case.
• Explaining precisely what the camera system captures, what the AI model detects, what data is retained and for how long, and who has access to it.
• Demonstrating the proportionality of the design — why a camera-based system is more appropriate than alternative safety measures, and why the specific camera placement and data scope are the minimum necessary.
• Committing to technical controls that prevent the system from being used for purposes beyond safety — and being prepared to discuss what those controls are.
• Establishing a review process by which the OR can assess the system's operation after deployment and raise concerns.

The OR consultation process is not an obstacle to good safety outcomes — it is a governance mechanism that, when engaged with seriously, produces more robust and legitimate deployments than those that bypass it. Crux Digits includes OR preparation support as part of our safety system implementation work, helping clients present the technical architecture in terms that enable informed consent decisions.

Privacy by Design: How to Build Safety Monitoring That Respects Workers

The phrase "privacy by design" has become something of a cliche in the AI industry, but its substance — that privacy protections should be built into the system architecture from the outset, not added as a compliance layer afterwards — is the right principle for factory safety monitoring.

The architectural choices that make privacy-by-design real in a computer vision safety system include:

Edge inference without central video storage. The most privacy-protective architecture runs the AI inference model directly on the camera or on a local compute unit at the edge of the network. The camera feed is processed locally; only the safety event signal (PPE alert, zone breach, proximity alert) is transmitted to the central system — not the video stream itself. This means there is no central video archive of worker activity to be misused, hacked or subjected to scope creep. This architecture is technically feasible with current edge AI hardware and is Crux Digits' default recommendation for factory safety deployments.

Anonymisation and aggregation at source. Where video does need to be retained — for incident investigation or model retraining — anonymisation techniques (blurring faces and identifying features, retaining only the safety-relevant event clip rather than continuous footage) reduce the personal data footprint substantially. Aggregate safety event statistics (number of PPE non-compliance events per zone per shift, proximity alert frequency by area) are almost always sufficient for the safety management purpose, and are far less privacy-sensitive than individual event records.

Role-based access controls. Access to any safety event data that could be used to identify individual workers should be restricted to the health and safety function, not available to line managers or HR as a general productivity monitoring resource. Access controls should be documented and audited.

Strict retention limits. Real-time alert systems need no video retention at all in normal operation. Incident clip retention should be limited to the period needed for investigation and any required regulatory reporting — not retained indefinitely as a general record of worker activity.

The EU AI Act and Worker Safety Systems in Manufacturing

The EU AI Act, which entered into force in 2024 and is being phased in through 2026 and beyond, has specific relevance to AI systems used in the workplace. Systems that monitor biometric data or that make decisions affecting workers in employment contexts receive particular attention under the Act. The EU AI Act text and the guidance published by the AI Office are the authoritative sources — this section summarises the general framework, not legal advice specific to your situation.

Most factory safety monitoring systems based on PPE detection and hazard zone monitoring are likely to fall into the limited or minimal risk categories under the EU AI Act, provided they are designed as alert systems that support human decision-making rather than automated systems that make consequential decisions about workers without human review. However, systems that use biometric categorisation, that perform real-time remote biometric identification, or that are used as a basis for employment decisions may face higher-risk classification and corresponding transparency and documentation obligations.

The key design principle for EU AI Act compliance in worker safety systems is that the system should assist and alert human safety managers — it should not replace human safety judgement or generate automated consequences for individual workers without human oversight. This is consistent with both good system design and the privacy proportionality requirements under GDPR.

Our AI implementation team includes regulatory compliance review as part of every safety system engagement, and our data engineering capability supports the data governance documentation that both GDPR and the EU AI Act require.

What a Responsible Deployment Looks Like: A Practical Checklist

The following checklist is intended for health and safety managers and operations directors evaluating or planning a computer vision worker safety system. It is not a legal compliance checklist — qualified legal and privacy counsel should be engaged for that purpose.

• Define the specific safety problem first. Which incidents are you trying to prevent? What is the incident rate for those event types? Where on the factory floor do they occur? The answers determine camera placement, detection scope and success metrics — and they are the evidence base for OR consultation and GDPR legitimate interest assessment.
• Choose the minimum monitoring scope necessary. Cover the specific hazard zones and entry points relevant to the safety problem, not the entire factory floor. Wider coverage requires stronger justification and creates larger data governance obligations.
• Design for edge inference without central video storage where technically feasible. This is the single most effective architectural choice for reducing privacy risk and enabling the proportionality argument under GDPR.
• Engage the OR before procurement, not after. OR consultation is a legal requirement, but it is also a practical necessity for a deployment that workers will accept as legitimate. Start the conversation early enough that OR feedback can genuinely influence system design.
• Document the GDPR legitimate interest assessment before go-live. This is a legal requirement under GDPR Article 6(1)(f) if legitimate interest is your processing basis, and it must reflect the actual system design, not a generic template.
• Establish technical controls that prevent use of the system for productivity monitoring or disciplinary purposes outside the safety remit. Write these controls into the system design, not just into policy documents.
• Communicate clearly with workers about the system. What is monitored, why, how data is used, who has access, and how long it is retained. Do this before go-live, not after workers have noticed new cameras.
• Build a review process. Who assesses system performance after deployment? When does the OR have an opportunity to raise concerns? What criteria would trigger a system review or discontinuation?
• Monitor for safety outcomes, not compliance rates. The success metric for a safety monitoring system is a reduction in workplace incidents, not an increase in PPE compliance scores. Track actual injury and near-miss rates against baseline, and be prepared to acknowledge if the system is not delivering safety improvements.

Ergonomics Monitoring: Opportunity and Limits

Musculoskeletal disorders are the most common occupational health condition in Dutch manufacturing, and ergonomic risk assessment is a statutory requirement under the Working Conditions Act. Computer vision systems can add genuine value here — but the specific use case matters enormously for the privacy analysis.

Workstation ergonomic analysis — where a camera is used to record a worker performing a specific task, the footage is analysed to generate a posture risk score, and the output is used to redesign the workstation — is a well-established and relatively uncontroversial application. The data collection is time-limited, the purpose is clear, the output is workstation improvement rather than individual worker assessment, and workers can give informed consent to participate. This use case maps well to the proportionality requirements under GDPR.

Continuous real-time AI ergonomics monitoring production of individual workers — where a camera monitors an individual's posture throughout their shift and the system generates alerts or reports on their ergonomic behaviour — raises substantially more significant privacy concerns. Even with safety intent, it is difficult to demonstrate that continuous individual monitoring is proportionate to the objective when workstation redesign (which removes the ergonomic risk for all workers performing the task) is available as an alternative. Crux Digits recommends workstation-level ergonomic analysis as the primary use case for most manufacturing environments, with continuous monitoring considered only where a specific risk profile justifies it after a proper proportionality assessment.

How Crux Digits Approaches Factory Safety System Deployments in the Netherlands

Crux Digits is a vendor-neutral AI consultancy based in Utrecht. We do not sell a proprietary camera platform or a packaged safety monitoring product — we design and build computer vision safety incident prevention systems tailored to the specific safety problems, physical environment and regulatory context of each client's factory.

For Dutch manufacturing clients, a typical safety system engagement covers four stages. In the first stage, we conduct a safety and site audit: reviewing your incident and near-miss records, identifying the specific hazard scenarios to be addressed, assessing camera placement options and lighting conditions, and mapping the regulatory and OR consultation requirements for your organisation. In the second stage, we design the system architecture — camera specifications, edge or on-premise inference setup, alert routing, data retention policy and access controls — and prepare the technical documentation needed for OR consultation and GDPR assessment. In the third stage, we build, test and deploy the system, starting with a pilot zone before full production rollout, and validate detection performance against agreed metrics. In the fourth stage, we support the OR review process, provide operator and safety manager training, and establish the monitoring and review cadence for ongoing system governance.

If you already have a camera infrastructure in place — conventional CCTV or a previous generation of machine vision — we can also conduct a standalone technical and governance audit, assessing whether your current setup is legally compliant and technically fit for the safety objectives you have identified.

Review our case studies for examples of AI safety and computer vision work in industrial environments, explore our engagement options, or get in touch to discuss your factory's specific safety challenges. For teams still at the evaluation stage, a no-obligation scoping conversation helps us understand your safety problem before recommending any technical approach.

For authoritative guidance on GDPR and workplace monitoring, the Autoriteit Persoonsgegevens publishes detailed guidance for Dutch employers on what is and is not permissible under Dutch data protection law. For the EU AI Act, the EU AI Act resource centre maintained by the Future of Life Institute provides accessible summaries of the Act's requirements alongside the official legislative text.

Frequently Asked Questions

Frequently asked questions