Detecting Worker Fatigue + Heat Stress with Vision (Saudi Summer Reality)

Why this matters in Saudi summer

Three structural realities make heat stress the dominant summer risk on KSA industrial sites:

1. Outdoor WBGT routinely exceeds 32°C from June through September.
2. PPE retention — workers cannot remove FRC, hard-hat or harness even when overheating.
3. Cultural factors — workers may underreport early symptoms to avoid being pulled off task.

The MoHRSD ban on outdoor work between 12:00 and 15:00 mitigates the worst of it, but does not eliminate the residual risk in early morning and late afternoon. Vision-based fatigue detection is the operational extension of the regulatory ban [VERIFY-SME — confirm current MoHRSD rule].

What vision can actually detect

Six markers are detectable with 2026 vision techniques:

1. Gait change — stride length shortening, asymmetry, drag.
2. Posture droop — shoulder rotation forward, head drop, back curve.
3. Micro-stops — repeated 5–15 second pauses outside normal task rhythm.
4. Repetitive head/face wipes — proxy for heat sweating beyond normal levels.
5. Sit-down events outside designated rest areas.
6. Slow recovery after lifting or carrying.

The underlying primitives are pose estimation and action recognition. For the foundational model context see DINOv3 and Grounding DINO.

The detection signal pipeline

A defensible 2026 pipeline:

1. Multi-frame pose estimation at 5–10 fps per worker track.
2. Gait extraction — stride length, cadence, asymmetry — over a 30-second rolling window.
3. Posture metrics — shoulder angle, head pitch — over the same window.
4. Environmental fusion — WBGT and ambient temperature from a co-located weather sensor.
5. Risk score — weighted combination, surfaced as Green/Amber/Red.

This is not a single-frame classifier. Heat stress is a temporal phenomenon; the pipeline must track over time.

Thresholds and escalation

A 2026-grade escalation path:

Risk score	Trigger	Action
Green	All markers normal	Log only
Amber	Two markers + WBGT > 30°C	Supervisor notified, water break recommended
Red	Three+ markers OR posture collapse	Immediate evacuation to cool zone

The escalation must be tied to the danger zone alerts solution and to the existing HSE radio chain. A vision alert that lives only on a dashboard does not save a worker.

Privacy and PDPL posture

Continuous behavioural monitoring of workers is a PDPL processing activity, and one with a higher sensitivity than asset tracking. The defensible posture:

1. Lawful basis — vital interest (worker safety) typically applies, but documented and reviewed annually.
2. Face-blurring at the edge for non-incident frames.
3. Retention — 14 days for non-incident clips, longer only when an incident is logged.
4. DPO sign-off with an explicit purpose statement.
5. Worker briefing — the workforce must understand the system exists and what it does.

For deeper context see the face recognition + PDPL piece and the PDPL compliance checklist.

False positives and the supervisor trust problem

Three controls bring nuisance alerts down:

1. Personal baseline. Each worker’s normal gait is recorded over the first three shifts; alerts fire when current gait diverges from personal baseline, not from a generic threshold.
2. Task-aware suppression. A worker performing a known cyclic task (digging, lifting) has different normal patterns than a worker walking; the system uses the active permit and zone to set context.
3. Environmental gating. Below WBGT 28°C, only the most severe markers trigger alerts.

Without these, the supervisor abandons the system within a week. See the hard-hat detection accuracy piece for the analogous PPE pattern.

Where the cameras go

Five rules for camera placement in fatigue/heat-stress monitoring:

1. Concourse and walkway cameras, not just task-zone cameras. Gait is most clearly visible during transit.
2. Rest-area cameras with privacy-respecting framing (no toilets, no prayer rooms).
3. Sun-shaded zones as the reference for “cool zone” evacuation.
4. Working-face cameras for direct posture observation.
5. Avoid pure overhead. Pose estimation from above loses too much information.

Cost envelope

Indicative SAR for a 200-worker industrial site:

Item	SAR per year
Vision software licence	90,000–180,000
Edge inference	45,000–80,000
WBGT sensor network	25,000–55,000
Integration with HSE radio	30,000–60,000 (Year 1)
Total Year 1	190,000–375,000

[VERIFY-SME for KSA-specific 2026 pricing.] The benchmark is workers saved from severe heat illness; conservative HSE costing typically values one prevented severe heat-stress incident at SAR 150,000–400,000 in lost time and medical cost [VERIFY-SME].

Common deployment mistakes

1. Single-frame classifier. Heat stress is temporal; single-frame detection is noise.
2. No personal baseline. Generic thresholds fire on anyone working hard.
3. No environmental fusion. Indoor air-conditioned zones get false alerts.
4. No HSE radio integration. Dashboard-only alerts fail to save workers.
5. No worker briefing — the system is perceived as covert surveillance and resisted.

Field deployment checklist

1. Camera placement covers concourses, rest areas and working faces.
2. WBGT sensors deployed and calibrated.
3. Personal baseline window of three shifts before live alerts.
4. HSE radio integration tested.
5. PDPL DPO sign-off and worker briefing complete.
6. Two-week shadow mode at start.

Next steps

If you are scoping fatigue or heat-stress vision detection on a Saudi industrial site, start with the danger zone alerts solution, the fall detection solution and the PPE detection solution. Cross-reference the Aramco EHS compliance guide for the SAP-PM hand-off and the face recognition + PDPL piece for privacy posture.

Book a heat-stress scoping call and we will produce a marker spec, threshold register and HSE escalation design within 10 working days.

React to this article

🔥...💡...🤔...👏...❤️...