Pular para o conteúdo principal
ACSMT — Assessoria e Consultoria Técnica
Back to blogStrategy

Computer vision, PPE, and behavior: ethics and regulatory compliance in the filmed factory

Cameras and vision models can reinforce PPE use and safe zones — with legal limits, bias, and clear policies on retention and transparency with workers.

Equipe ACSMT

Equipe ACSMT

Time Técnico

2 min read
Surveillance and security cameras in an urban industrial environment

Executive TL;DR

Computer vision applies neural networks to video streams to detect helmets, goggles, or proximity to a moving machine. In OHS, this can accelerate feedback and reduce human observer bias, but it introduces psychosocial risk of surveillance and discrimination if poorly designed. LGPD requires a legal basis, data minimization, and transparency; NR-12 requires that interlocks do not allow startup without protection, regardless of AI. The PGR must list intensive monitoring as a psychological risk agent.

Table of contents

  1. Safe use cases in cells and logistics
  2. Engineering: accuracy, lighting, and drift
  3. Human governance: HITL and audit
  4. Privacy, CIPA, and unions
  5. Integration with NR-12, NR-10, and eSocial
  6. Typical error: algorithmic paternalism
  7. FAQ
  8. References

Safe use cases in cells and logistics

A zone maintained by laser or light curtain fails less when a camera confirms human presence before reset — if the logic is dual-channel. For forklifts, collision alerts help — but they do not replace traffic design.

Engineering: accuracy, lighting, and drift

A model trained in bright light fails at night with only blue LED. Retrain per line. A PPE macro is worthless if legal dark goggles exist.

Human governance: HITL and audit

Define when an alarm only opens a ticket and when it stops the machine. Human-in-the-loop reduces serious false positives in accidents.

Privacy, CIPA, and unions

Public camera map for workers. Avoid faces in datasets — automatic blur. Short retention if there is no incident.

Integration with NR-12, NR-10, and eSocial

A filmed accident requires a chain of custody for S-2210 and labor lawsuits. NR-10: electric shock does not wait for AI.

Typical error: algorithmic paternalism

Disciplining by algorithm without union framework and documented just cause creates massive liability.

FAQ

Does AI replace CIPA?

No; it complements inspection.

Can I use facial recognition?

High sensitivity — avoid in large areas without ANPD analysis.

What about offshore cloud?

SCC clause + local law.

References

  1. Brazil. LGPD and ANPD — AI and biometrics.
  2. Brazil. NR-1, NR-12, NR-10.
  3. ISO/IEC 42001 — AI management (emerging international reference).

Editorial note: AI legislation is in progress — review annually.

Related reading

Keep reading

Team meeting in an office environment — work tension and pressure
NR-1

Burnout, moral harassment, and RMP: mapping risk without denying worker dignity

Burnout and harassment require method in the NR-1 inventory: safe channels, impartial investigation, and measures that change management — beyond brochures. LGPD and union dialogue in the same package.

Equipe ACSMT
6 min read
Executive meeting in a modern room with an urban view
Strategy

Enterprise OHS consulting: criteria for directors, COO, and risk committee

Choosing an OHS partner through a mature RFP: methodology, evidence, PGR–eSocial integration, and decision-making that does not push responsibility onto the SESMT alone.

Equipe ACSMT
2 min read
Workers wearing safety helmets in an industrial environment
Strategy

Safety culture, BBS, and human factors: fair feedback that fixes conditions first

Behavior Based Safety works when there is organizational justice, parity in observation, and investment in engineering — not just posters and observation quotas.

Equipe ACSMT
2 min read