Safety Norms: Computer Vision Systems

1Purpose, Scope, and Safety Principles

These norms set the minimum safe and compliant practices for computer vision work. Computer vision means systems that capture images or video from cameras, sensors, or other imaging devices and use models to detect, classify, count, track, measure, or describe what they see. The norms cover laboratory prototypes, field pilots, production deployments, and retired systems, along with the people, cameras, edge devices, servers, datasets, model files, and dashboards involved at each stage.

The guiding principles are to collect only what is needed, keep humans accountable for consequential decisions, protect people in view, secure every connected component, and treat model output as a probability rather than a fact. Where a rule depends on jurisdiction, such as biometric processing, workplace monitoring, public surveillance, or data protection, this document labels it as jurisdiction dependent. Owners must confirm the current legal position with competent authorities or qualified counsel before deployment.

• Mandatory: every system must have a written purpose, a named owner, and a retention rule before it captures live imagery.
• Mandatory: this document does not replace law, vendor instructions, or approved organizational policy; where they conflict, the stricter requirement applies.
• Recommendation: prefer the least intrusive sensing method that meets the operational goal, such as counting zone occupancy instead of identifying individuals.

2Roles, Competence, and Accountability

Each deployed system needs one accountable owner who approves purpose, data sources, risk rating, and go-live decisions. Owners are responsible for the system record, periodic reviews, and retirement. Responsibility cannot be handed to a vendor, although vendors may perform defined tasks under a written contract.

Engineers build, label, train, and test models. Security staff review network placement, credentials, and access design. Privacy or compliance staff assess personal data use and notice requirements. Site operators manage physical installation, lens cleaning, and status checks. Where one person holds several roles, document a compensating review so that no single person can deploy, approve, and audit their own work.

• Mandatory: maintain a current register of cameras, models, datasets, vendors, and administrator accounts, reviewed at the interval set by the risk rating.
• Mandatory: staff who access live or archived video containing people must complete privacy and security training before access is granted.
• Mandatory: only trained and authorized installers may mount cameras at height or connect them to building power, following local electrical, building, and work-at-height rules and the site safety program.
• Recommendation: require vendors to supply model limitation documents, patch commitments, and an end-of-support plan before purchase.

3Course-Specific Hazards and Risk Assessment

Computer vision hazards fall into five groups. Privacy harm arises when identifiable people, license plates, screens, or documents are captured or retained. Security harm arises when cameras, video streams, or model files are compromised. Physical safety harm arises when a missed or wrong detection influences machinery, vehicles, or people. Reliability harm arises when performance drops under new lighting, weather, or camera movement. Misuse harm arises when a system is used for a purpose nobody approved.

Rate each hazard as low, medium, or high using likelihood and severity scores agreed by the owner and the safety or privacy lead. Any high rating requires documented mitigation and sign-off before go-live. Re-rate whenever purpose, location, the population observed, or the model changes.

• Privacy: faces, plates, and sensitive screens in frame; treat as high unless the design masks or minimizes them before storage.
• Security: default passwords on network cameras, unencrypted streams, exposed administrative interfaces, and tampered model weights.
• Adversarial manipulation: printed patches, stickers, or altered signs that cause misclassification; assess wherever results trigger access or safety actions.
• Physical: missed people in robot cells, forklift lanes, or loading docks; a vision output alone must never be the only protection for hazardous motion.
• Reliability: false positives from shadows, glare, rain, or bumped cameras; false negatives on rare objects or poses absent from training data.
• Misuse: repurposing an occupancy counter to track individual employees or customers without a new approval.

4Preventive Controls and Safe Work Practices

Build controls into the design. Collect only the imagery required for the stated purpose, reduce resolution or crop frames when detail is unnecessary, and mask or blur faces and plates where the task does not need them. Process data at the edge when feasible so raw video does not leave the site. Record the source and legal basis of every training image, along with labeling instructions and dataset version identifiers.

Test before release with data that reflects real conditions: each camera angle, lighting band, weather condition, and product or population the system will encounter. Define minimum acceptable performance and documented alert thresholds before go-live. Treat any safety-related function as one layer in a wider safety design, never as a replacement for guarding, interlocks, or trained human checks.

• Mandatory: keep development, test, and production environments separate; do not train on production video without an approved data use record.
• Mandatory: do not treat vision output as the sole basis for an action affecting physical safety, access control, or employment without defined human review.
• Mandatory: live tests against cameras, networks, or third-party sites require written authorization from the system owner and the network or site owner, with scope, time window, and rollback steps written first.
• Mandatory: adversarial and penetration testing only in isolated lab environments or on systems covered by a signed test authorization.
• Recommendation: review confidence thresholds with operators before any change, and log the reason for each change.
• Recommendation: sample predictions by camera and condition at a regular interval set by the risk rating.

5Equipment, Environment, Information, and Protective Safeguards

Hardware and network safeguards protect the cameras and the people they observe. Mount equipment according to manufacturer instructions, and confirm that enclosures, cables, and power supplies suit the site environment. Place cameras so they do not capture restrooms, changing areas, private residences, or restricted zones unless a documented, lawful reason and notice exist. Clean lenses on a scheduled basis and aim fixtures to avoid glare from reflective surfaces.

Information safeguards cover credentials, transport, storage, and access. Use unique accounts, multifactor authentication for administrators, and role-based permissions for live view, archives, and model management. Encrypt video and model files in transit and at rest where the platform supports it. Hazardous motion must be protected by hard guarding and approved safety-rated controls, with vision added only as an extra layer.

• Mandatory: change factory default passwords before connection and disable unused ports, protocols, and remote services.
• Mandatory: install firmware only from verified vendor sources, following a patch schedule approved by the owner.
• Mandatory: post visible notices where local law requires recording notice, and keep the legal basis in the system record; verify the requirement locally.
• Mandatory: limit footage retention to the approved period, delete it automatically, and log every export.
• Recommendation: place cameras and processing units on a segmented network with no direct internet exposure unless a security review approves it.
• Protective equipment: installers working at height or with power tools use the equipment specified by the site safety program and must not work alone on elevated or energized installations.

6Incident Reporting, Emergency Response, and Escalation

Report any incident to the system owner and security contact as soon as it is noticed. Incidents include unauthorized access to video, a camera found pointing at a restricted area, suspected tampering, a model failure that caused an operational error, and near misses where the system almost caused harm. Near misses deserve the same attention because they reveal weak controls before injury or exposure occurs.

Responders contain first and investigate second. Preserve logs, configuration snapshots, and relevant footage under the legal hold process before wiping or reimaging devices. Escalate suspected privacy breaches to the privacy lead immediately, who will assess notification duties under the applicable law with counsel. Do not attempt to recover data from unauthorized sources.

• Isolate a compromised camera or disable the affected alert path, and document every change made.
• For a physical hazard, stop equipment through its approved safety system; never rely on a vision alert to stop machinery.
• If someone is injured or in immediate danger, contact local emergency services before the system owner.
• Notify the privacy lead and security contact of any suspected exposure of personal images; notification duties and recipients depend on local law, so verify them with counsel.
• Hold a blameless review with documented corrective actions, named owners, and verification steps.

7Responsible Use and Compliance

Responsible use asks whether the system should be deployed at all, not only whether it can be. Before approval, the owner documents the benefit, the people affected, alternatives considered, and why the chosen design is proportionate. Biometric identification, emotion inference, and public or workplace surveillance carry specific legal and ethical constraints that are jurisdiction dependent and must be checked with qualified professionals before design begins.

Communicate limitations to users. Operators should know what the model cannot reliably do, such as identify people in low light or crowded scenes, and should be trained to override or escalate. Do not describe the system as more accurate, neutral, or secure than tests support.

• Mandatory: no biometric identification, emotion inference, or covert monitoring without a documented legal review and approval from the owner and privacy lead.
• Mandatory: do not reuse data or models for a new purpose without a new assessment.
• Recommendation: publish a plain-language summary of what the system detects, what it stores, and how people can ask questions.
• Recommendation: include affected workers or community representatives in design reviews where practical.

8Records, Training, Monitoring, and Review

Documentation proves that controls exist and supports investigation. Each deployment needs a system record containing purpose, owner, risk rating, data flows, retention periods, model and firmware versions, test results, known limitations, and sign-offs. Keep change logs for model updates, threshold adjustments, and camera relocations. Retain records as long as local policy requires.

Training is role based and refreshed when the system, applicable law, threat picture, or site changes. Monitoring checks accuracy against the test baseline, drift, camera health, access logs, and unusual exports. Full reviews occur at intervals set by the risk rating and after every incident. Lessons feed back into these norms and local procedures.

• Monitor false positive and false negative rates, and investigate any sustained shift from the approved baseline.
• Review access logs for off-hours queries, bulk downloads, or new administrator accounts.
• Repeat privacy and security assessments after major changes in purpose, location, camera placement, or model.
• Decommission systems by removing stored footage, personal data in model artifacts, and credentials, then record completion.
• Verify that retention and records rules match current requirements with the competent authority or a qualified professional.

Implementation checklist

• Confirm a written purpose, named owner, and retention rule exist before any live capture begins.
• Change default credentials and disable unused services on every camera and edge device.
• Assess privacy, security, physical safety, reliability, and misuse risks, and record the ratings.
• Mask or minimize identifiable data, and confirm the legal basis and notices with qualified staff.
• Test with representative conditions and set documented alert thresholds before release.
• Obtain written authorization before any live test, network scan, or adversarial evaluation.
• Keep vision output out of sole control of hazardous motion or consequential decisions.
• Restrict video and model access by role, log all exports, and enforce automatic deletion.
• Report incidents and near misses immediately, isolate compromised devices, and preserve logs.
• Review accuracy, access logs, and risk ratings on schedule and after every significant change.