Site Safety Computer Vision (PPE, Zones, Vehicles)

What problem does this solve?

Most industrial sites already have cameras everywhere and almost no safety benefit from them. The recordings are used after an incident, to establish what happened and who was at fault. Nobody is watching sixty streams at once, so the near-misses that precede a serious injury — the person walking through a forklift aisle, the contractor entering a live lifting zone, the crew working at height without harnesses because the harnesses were in the other container — are never seen unless a supervisor happens to be standing there.

The controls that exist depend on human vigilance at exactly the moments when vigilance is weakest: end of shift, during a rush, when a subcontractor crew is on site for two days and has not absorbed the site rules. HSE teams respond with more toolbox talks and more spot checks, which produce compliance while the observer is present. The result is a safety record built on lagging indicators — incidents and lost time — with no measurement of the unsafe conditions that produced them.

There is also a credibility trap. Sites that buy an off-the-shelf 'AI safety camera' product often discover within weeks that it flags a rolled-up jacket as a missing vest, misses helmets on people twenty metres from the camera, alerts on the same stationary worker forty times an hour, and cannot tell a supervisor crossing a marked line from a labourer standing under a suspended load. The alerts go to a WhatsApp group, the group mutes it, and the system becomes an expensive recorder again. The failure is rarely the model alone — it is camera views the detections were never viable at, zones drawn without operational input, and no policy for what happens when an alert fires.

How the solution works

Safety computer vision works when it is engineered as a site system rather than bought as a feature. Existing ONVIF/RTSP cameras are surveyed view by view to establish what is actually resolvable: a person occupying only twenty pixels of height cannot be assessed for PPE, no matter what the model claims. Cameras that can support a detection are assigned to it; those that cannot are either repositioned, given different optics, or used only for coarse detections like presence in a zone. This detection matrix — camera by camera, event by event — is the deliverable that keeps expectations honest before anything is installed.

Analytics run on an on-premise node so video never leaves the site. Person and vehicle detection, PPE attribute checks, multi-frame tracking and zone or distance rules produce events; an alert policy then decides which events reach a human, with severity, confidence thresholds, cooldown periods and shift-based routing. Alerts go where they can change the outcome: a local beacon or horn near the hazard, a message to the supervisor's phone, an entry in the control room queue, and a clip in the HSE dashboard for trend analysis. Swedish Technology builds these systems as custom development on the customer's own infrastructure, and can integrate them with the site's existing surveillance platform, including our own RAQEEB AI surveillance and warehouse safety work.

1. 1
   Input Site layout, camera positions and specifications, the list of hazards and rules to enforce, shift patterns, and who is expected to respond to each type of alert.
2. 2
   Capture Existing cameras are read over RTSP/ONVIF, or new cameras are placed where the survey shows a viable view. Streams are decoded on the analytics node; frame rate and resolution per stream are set to what the detection needs, not to the maximum.
3. 3
   Detection Models detect people, vehicles (forklifts, trucks, plant) and PPE attributes such as helmet, high-visibility vest, and where the view permits, harness or hearing protection. Detections are tracked across frames so one person is one object, not a new alert every frame.
4. 4
   Zone & rule evaluation Calibrated zones drawn on each camera view — exclusion areas, PPE-required areas, pedestrian walkways, vehicle aisles — are evaluated together with dwell time and, where geometry allows, distance between a vehicle and a person.
5. 5
   Alert policy Events are filtered by confidence, minimum duration and repeat suppression, assigned a severity, and routed by zone and shift. A person crossing a walkway for two seconds is not the same event as someone standing in a lifting zone for a minute.
6. 6
   Action High-severity events trigger local warning devices near the hazard and a supervisor notification; medium events enter a review queue; all events are stored with a short video clip as evidence.
7. 7
   Reporting & learning Trends by zone, shift, contractor and event type feed the HSE review, and every dismissed alert is recorded with a reason so thresholds, zones and models can be improved with evidence.

Reference architecture

Video stays on site, decisions are auditable, and nothing in the safety-critical chain depends on this system. Five layers, deliberately kept separate.

Deployment options: Standard deployment is fully on-premise: analytics node in the site server room or a ruggedised cabinet, storage local, no outbound internet required. Multi-site customers can add a private-cloud reporting layer in the UAE that receives aggregated event metadata only, with video remaining at each site unless explicitly exported.

Key capabilities

Integrations

Safety vision should reuse the site's existing surveillance and operational systems rather than creating a parallel stack. All integrations are engineered against the customer's own platforms.

Industry use cases

Implementation approach

1. 1
   Hazard and rule workshop (1 week) With HSE and operations, list the incidents and near-misses actually being targeted, the rules to enforce, the exemptions that exist in practice, and who will respond to each alert type.
2. 2
   Camera survey and detection matrix Assess each camera view for pixels-on-target, angle, occlusion and lighting; produce a written matrix of which detections are viable per camera, which need repositioning or new optics, and which are not achievable.
3. 3
   Zone design and calibration Draw zones on the camera views with the people who work in them; calibrate the ground plane where distance-based rules such as forklift proximity are needed.
4. 4
   Pilot on a limited camera set (4–8 weeks) Run detections in silent mode on a handful of representative views, measure detection and false-alarm rates against reality, and tune thresholds before anyone receives an alert.
5. 5
   Alert policy and response design Agree severity levels, routing by shift and zone, cooldown periods, escalation, and the maximum number of alerts per shift the team commits to reviewing.
6. 6
   Integration build VMS event push, local warning devices, supervisor app or control-room queue, HSE dashboard and export to the incident management system.
7. 7
   Roll-out and training Extend to the remaining viable cameras, train supervisors and operators in Arabic and English, and brief the workforce on what the system does and does not do — quietly installed safety cameras generate resistance that undermines the programme.
8. 8
   Review cycle and hand-over Monthly review of alert precision, zone changes as the site layout evolves (essential on construction sites), retraining schedule, documentation and source hand-over.

Security & deployment

The analytics node sits on the site's camera network segment, reads streams read-only, and requires no outbound internet connection; where interlocks with barriers or machinery are requested, they are implemented only after the customer's own safety assessment and never as a substitute for a safety-rated device. Access to live views, clips and event history is role-based and logged, so it is possible to demonstrate exactly who viewed which footage. Retention is short by default and configurable, face blurring can be applied at ingest so unblurred frames are never stored, and event exports carry metadata rather than video unless explicitly authorised. For classified or defence-adjacent facilities the full system runs air-gapped, with model updates delivered as signed offline packages.

Limitations & prerequisites

• This is a warning and evidence layer, not a safety-rated control system. It does not replace light curtains, guarding, interlocks, proximity sensors on vehicles or a permit-to-work regime, and it must never be presented in a risk assessment as if it did.
• Detection quality is governed by the camera view. Distance, angle, occlusion by racking or equipment, backlight, glare, rain, dust and night lighting all reduce accuracy, and some views simply cannot support PPE attribute checks at all.
• PPE detection is appearance-based. A helmet carried in the hand, a vest under a jacket, a non-standard colour or an unusual harness type will produce errors in both directions; correct wearing (chin strap fastened, harness clipped) is generally not verifiable from a camera.
• Distance-based rules such as forklift-person proximity depend on ground-plane calibration and degrade with camera movement, changed layouts or unusual perspectives; they give a useful warning zone, not a measured safety distance.
• False alerts cannot be eliminated. The system is tuned to a level of alerts the team agrees to handle, which necessarily means some events are missed — that trade-off must be documented and accepted by the HSE owner.
• Sites change. On construction projects zones must be redrawn as work progresses, and an unmaintained zone map produces alerts that everyone learns to ignore within two weeks.
• Worker consent, privacy regulation and industrial-relations context apply. Deployments that are not communicated to the workforce reliably attract obstruction — cameras turned, blocked or reported as faulty.
• Identifying individuals from safety alerts is deliberately out of scope in our designs unless the customer has a documented legal basis; reporting is by zone, shift and contractor.

Supervisor observation vs sensors vs safety vision vs safety-rated systems

These layers do different jobs. Safety vision fills the gap between human observation and hard-wired protection, and does not replace either.

A practical site combines them: safety-rated protection at machines, tag-based proximity on high-risk vehicles, vision for site-wide coverage of untagged people and areas, and supervisors focused on the exceptions the system surfaces.

FAQ

Related topics & solutions

• Industrial Vision Inspection with AI
• Predictive Maintenance AI for Industrial Assets
• Pipeline & Corrosion Vision Inspection AI
• UWB & BLE RTLS for Indoor Positioning
• Smart Warehouse Safety
• RAQEEB ÔÇô AI Surveillance
• Octopus WMS
• MOWQIE ÔÇô RTLS Tracking