In brief

  • The problem: many AI safety camera trials start with hardware and end with a vague result, because nobody defined the hazard, the baseline or what success looks like.
  • The requirement: Australian WHS law expects you to work through the hierarchy of controls, consult workers before changes, and, in New South Wales, give written notice before camera surveillance starts.
  • The solution: a site readiness checklist that prepares the hazard, the people, the cameras and the measures before a Flagman trial begins.
  • The outcome: a trial that produces evidence you can act on, operators who trust the alerts, and a clear decision on whether to scale.

The cameras arrive on a Monday. By Wednesday they are mounted, by Friday the alerts are flowing, and by the third week the operators have learned to ignore them. When the trial ends, the safety manager has a folder of clips and no clear answer for the operations executive who asked whether it worked.

That outcome is common, and it is rarely the technology’s fault. A trial is a test of a control in a real workplace. If the site is not ready, the test cannot tell you much. This AI safety camera trial checklist sets out what to settle before the first camera goes up.

The problem: trials that cannot answer the question

The risk that buyers of AI safety cameras in Australia are trying to reduce is serious. Safe Work Australia reports 188 worker fatalities in 2024, a rate of 1.3 per 100,000 workers. At least one vehicle was involved in 66 per cent of those deaths (124 workers). Transport, postal and warehousing recorded 54 fatalities, and mining recorded 10, a rate of 3.4 per 100,000 workers.

Regulators are pointing sites toward technology as part of the answer. WorkSafe Victoria reports eight forklift-related deaths since 2019, three of them pedestrians, and suggests employers consider proximity devices that detect pedestrians and slow mobile plant. In May 2026, SafeWork NSW published a review of technologies for mobile plant that recommends consultation, trial periods, and treating technology as an addition to existing controls rather than a replacement.

With that pressure, it is tempting to run a quick pilot. The trouble is that a rushed trial tends to fail in predictable ways:

  • The hazard is described as “safety” rather than a specific interaction, such as forklifts and pedestrians at one crossing.
  • There is no baseline, so nobody can compare before and after.
  • Workers hear about the cameras after they are installed, and trust is lost early.
  • Alert zones are left at default settings, operators receive alerts they see as false, and alert fatigue sets in.
  • The trial is treated as a replacement for barriers or procedures rather than a layer on top of them.

The requirement: what Australian law expects before you start

An AI safety camera trial changes how a workplace operates, so the usual WHS duties apply. Name your jurisdiction: most states and territories use the model WHS laws, while Victoria uses the Occupational Health and Safety Act 2004.

  • Work through the hierarchy of controls. Safe Work Australia’s model Code of Practice, How to manage work health and safety risks, says eliminating the hazard is the most effective control. Administrative controls and personal protective equipment rely on human behaviour and supervision and tend to be least effective on their own. A camera alert must sit inside that hierarchy, not in place of higher controls.
  • Consult workers and representatives. Safe Work Australia states that consultation is a legal requirement, including when proposing changes that might affect workers’ health and safety. Health and safety representatives must be consulted too.
  • Follow workplace surveillance law. In New South Wales, the Workplace Surveillance Act 2005 requires written notice to employees at least 14 days before surveillance starts, unless an employee agrees to a shorter period. The notice must state the kind of surveillance, how it will be done, when it starts, and whether it is continuous or intermittent. Cameras must be clearly visible, with signs at each entrance to the area.

For mines, the Queensland Mines Inspectorate’s Guidance Note 27 on collision prevention is a useful reference. It says collision avoidance systems, including cameras, cannot replace mine design, engineering and administrative controls. It also asks how the system handles false alarms and how its effectiveness will be measured.

Why existing controls leave a gap

Most sites already run traffic management plans, guarding, isolation procedures and supervision. These controls are essential, but they share one weakness: they do not tell you when they are being bypassed. A painted walkway does not report that someone stepped outside it, and no supervisor can watch every aisle on every shift. Near misses go unreported, so the plan is reviewed against what people remember rather than what happened.

Computer vision workplace safety tools are designed to fill that gap: detection at the point of risk, and data on how the controls perform. A well-prepared trial is how you find out whether they fill it at your site.

The solution: how Flagman supports a structured trial

Flagman uses AI vision to detect people, vehicles and personal protective equipment in camera footage. Processing happens on site at the edge, with no cloud or internet connection required. It works with ONVIF IP, USB and GMSL cameras, so many sites can trial it on existing CCTV before adding new cameras.

Features that make a trial measurable include:

  • Configurable detection zones. Zones are drawn around the hazard being tested, such as a pedestrian crossing, a crane load path or a conveyor transfer point. In a typical forklift setup, a warning is raised at about 2.5 metres and a critical alert at about 1.5 metres.
  • Time-stamped video event logs. Every detection is recorded with video to a central dashboard, so the safety team can separate real interactions from nuisance alerts.
  • Integration outputs. Relay outputs and programmable logic controller integration over Modbus and PROFINET can drive warning lights or signals. Where a machine’s controls allow, Flagman can send a slow or stop signal.
  • Privacy by default. The default Australian configuration is people, vehicle and PPE detection. Facial recognition is an optional module, not part of a standard trial.

Limitations: what a Flagman trial will not prove

  • Flagman does not replace elimination, separation, physical guarding, isolation or lockout. It adds detection and evidence on top of those controls.
  • Flagman has no published Safety Integrity Level or Performance Level rating, and it is not a certified presence sensing safeguarding device. Its slow or stop output is not a certified safety function.
  • Flagman has no hazardous area certification and must not be trialled in underground coal or explosive atmospheres.
  • Detection depends on camera position, lighting, dust, weather and line of sight. A short trial in one area does not prove performance across a whole site.
  • Camera data is regulated. Even with on-site processing, privacy and surveillance obligations still apply.

The AI safety camera trial checklist

  1. Name one hazard and one area. For example, forklift and pedestrian interaction at the dispatch crossing, or workers under suspended loads in bay three.
  2. Confirm higher controls are in place. Check separation, barriers, guarding and isolation first, and record any gaps the trial is meant to cover.
  3. Set a baseline. Gather two to four weeks of incident and near miss records for the area, and decide what the trial will compare against.
  4. Define success in writing. Agree measures before go-live, such as interactions detected per shift, alerts reviewed, nuisance alerts removed and corrective actions closed.
  5. Consult early. Brief workers and health and safety representatives on what is detected, what is recorded, who can view footage and how long it is kept.
  6. Give surveillance notice. In New South Wales, issue written notice at least 14 days before the trial starts, fit signs at each entrance and keep cameras visible. Check the rules in your own state or territory.
  7. Survey the cameras and network. Confirm existing CCTV angles, lighting, network capacity and security, and add cameras only where coverage falls short.
  8. Tune before you judge. Run the first one to two weeks as a tuning period, adjusting zones with operators. This is where alert fatigue in safety systems is prevented, before any results are counted.
  9. Plan maintenance. Schedule lens cleaning, camera health checks and recalibration, especially in dusty or washdown areas.
  10. Review weekly, then decide with evidence. Look at events by shift and location, record what changed, and compare results against the baseline before you scale, adjust or stop.

If you are still choosing between technologies, read our scorecard for choosing a pedestrian detection system. For privacy settings, see our guide to deploying AI safety cameras lawfully.

Keeping the system healthy: maintenance after the trial

A detection system is only as reliable as its weakest camera. Dust, steam, vibration, a bumped mounting bracket or a network fault can quietly reduce coverage, and nobody notices until an event is missed. SafeWork NSW’s technology review makes the same point: keep up calibration and maintenance, and check the IT infrastructure behind the system. Treating camera health, network uptime and alert volumes as maintenance data lets sites predict and fix problems before coverage drops.

Flagman in Australia: local delivery by Sharktech Global

Sharktech Global, based in Liverpool, New South Wales, is the exclusive Australian and New Zealand delivery partner for Perleybrook Labs, the developer of Flagman. For a trial, Sharktech runs the site assessment, provides a written, costed plan before any commitment, installs and tunes Flagman, and supports it locally. Because cameras sit on site networks, Sharktech also provides managed IT, Essential Eight cybersecurity assessments and advice on separating office, camera and production systems. Details are on its pages for warehouses, manufacturers and mining operations.

The outcome you can measure

  • A documented baseline and a like-for-like comparison at the end of the trial.
  • Alerts delivered at the point of risk, with nuisance alerts identified and reduced during tuning.
  • Near misses captured as time-stamped data, by area, shift and time of day.
  • Consultation and notice records that show the trial was run lawfully.
  • A clear, evidence-based recommendation for your board on whether to scale.

Key takeaways

  • A trial tests a control, so define the hazard, baseline and success measures first.
  • Keep AI safety cameras inside the hierarchy of controls, as a layer on top of separation, guarding and isolation.
  • Consult workers and follow workplace surveillance law before the cameras go live.
  • Tune alert zones with operators, maintain the cameras, then judge the results.

Next step

Request a Flagman safety assessment to scope a trial area, review your existing cameras and agree the measures before you start, or call Sharktech Global on 0468 017 373.

Sources: Safe Work Australia, Key WHS Statistics Australia 2025; Safe Work Australia, Model Code of Practice: How to manage work health and safety risks; Safe Work Australia, Consultation; WorkSafe Victoria, New guidance and inspections target forklift safety (22 May 2024); SafeWork NSW, Technologies for mobile plant (May 2026); Workplace Surveillance Act 2005 (NSW), sections 10 and 11, as passed by Parliament; Queensland Mines Inspectorate, QGN 27 Collision prevention (December 2017). This article is general information only and is not legal advice. Seek advice specific to your site and jurisdiction.