Real case: a mid-sized Mexican construction company (3 active sites simultaneously, ~250 workers at peak) deployed a computer vision system for automatic PPE (Personal Protective Equipment) detection on a vertical construction site in Mexico City. 12 IP cameras + edge server with Jetson Orin + fine-tuned YOLOv8 model. Here are the numbers: real accuracy, false positives, cost, ROI, and lessons learned.
( anonymized data — company requested no name publication, but proportions and metrics are shareable )
The problem
Construction is the #1 industry for fatalities in Mexico and the US:
- In Mexico, the construction accident rate is ~28% higher than the industrial average per IMSS 2023 data.
- In the US, OSHA reports 1,056 fatal injuries in construction in 2022 (20% of all industrial fatal injuries) [BLS workplace injuries].
- 85% of fatal injuries in construction involve some form of PPE failure (no hardhat, no vest, harness improperly worn).
The classic operational problem: the safety supervisor spends 4 hours/day on rounds verifying that workers wear hardhat, vest and glasses. Even so, on a site with 80+ active workers, there is a margin where 5-10% of personnel are without PPE at any moment.
System design (Q1 2026)
Technology stack:
- 12 4K IP cameras Hanwha XNV-9083R (vandal dome IP67) — resistant to site dust
- 3 NVIDIA Jetson Orin AGX edge servers (one per site zone: foundation, structure, finishes)
- YOLOv8m fine-tuned model on 8,000 annotated images of the specific site
- Detected classes: hardhat, reflective vest, safety glasses, harness (in height zones), person without PPE
- Real-time dashboard (Grafana + InfluxDB) accessible to safety supervisor and site manager
- Alerts: snapshot + local sound when person without PPE crosses designated zone
CAPEX breakdown:
| Component | Quantity | Unit cost (USD) | Subtotal (USD) |
|---|---|---|---|
| Hanwha XNV-9083R 4K camera | 12 | $480 | $5,760 |
| NVIDIA Jetson Orin AGX 64GB | 3 | $1,999 | $5,997 |
| Industrial PoE+ switch (8 ports) | 3 | $320 | $960 |
| Industrial Cat6 cabling + installation | 12 cameras | $180/cam | $2,160 |
| Dashboard server (x86 mini PC) | 1 | $850 | $850 |
| SSD storage 4TB (90 days of alerts) | 1 | $380 | $380 |
| Hardware subtotal | — | — | $16,107 USD |
| Software (YOLOv8 fine-tuning + integration) | 1 project | — | $8,000 |
| Installation + commissioning | 1 | — | $3,500 |
| Total CAPEX | — | — | $27,607 USD (~MXN $550,000 at Q1 2026 exchange rate) |
Annual OPEX:
- Maintenance (camera cleaning, model retraining, support): ~$3,500 USD/year
- Additional electricity (3 Jetson + switch): ~$720 USD/year
- Year 1 OPEX: ~$4,220 USD (~MXN $84,000)
Operational metrics (6 months post-deployment, Q2-Q3 2026)
Model accuracy:
| Class | Precision | Recall | F1-score |
|---|---|---|---|
| Hardhat | 0.94 | 0.91 | 0.92 |
| Reflective vest | 0.92 | 0.88 | 0.90 |
| Safety glasses | 0.86 | 0.79 | 0.82 |
| Harness (in height) | 0.89 | 0.85 | 0.87 |
| Person without PPE (alert) | 0.91 | 0.84 | 0.87 |
The model achieves 87% F1 on the critical class (no-PPE alert). The 13% error breaks down as:
- False negatives (missed real infraction): 16% — person with back to camera, person partially occluded by material, person >25 m from camera
- False positives (alert for non-existent infraction): 9% — confusion with hardhats on the floor, shadows, similar visuals
Alert rate per day:
- Average: 38 alerts/day
- Valid alerts (real infraction): ~28/day
- False alarms: ~10/day (26%)
- Review time per alert: ~30 seconds (fast visual confirmation)
Operational adoption:
- Before system: 1 in-person round every 4 hours = 6 rounds/day × 30 min/round = 3 person-hours/day
- After: 6 rounds + review of 38 alerts = 4 person-hours/day (more coverage, but with added desk work)
- Net time added: 1 person-hour/day for safety supervisor
But the real change is in compliance rate:
- Pre-system: ~85% of personnel with full PPE at any moment
- Post-system (month 3 onwards): ~96% of personnel with full PPE (people know there are cameras and automatic alert)
- Absolute improvement: 11 percentage points = 9 more people with full PPE at any moment
Real ROI
Avoided cost per incident:
- A minor injury in Mexican construction costs ~$45,000 MXN (medical care + lost days + risk premium)
- A major injury (fracture, fall) costs ~$280,000 MXN
- A fatal injury costs ~$3.5M MXN (compensation + sanctions + risk premium + site shutdown)
- In 6 months, the site had 0 fatal injuries (vs 1 fatal injury in the same period of the previous year, pre-system)
- Avoided minor injuries estimate: 4 (vs 7 in equivalent prior period)
- Estimated gross savings: 3 avoided minor injuries × $45,000 = $135,000 MXN (we do not assume credit for the avoided fatal injury because it is statistically debatable with n=1)
Intangible benefit:
- Risk premium reduction on policy renewal: ~$120,000 MXN/year (estimated, based on industry benchmark)
- Compliance with contractual clause with client (constructor requires verified PPE compliance): renewed for 1 additional year = ~$8M MXN of contract retained
Estimated annual benefit: ~$255,000 MXN tangible + $8M MXN intangible (contract retention).
Payback period: CAPEX $550,000 MXN / tangible benefit $255,000 = ~26 months on tangible benefit alone. If contract retention is included, payback <1 month.
The 3 things that went wrong
1. Dust and vibration on site degrade cameras Cameras were mounted on site with heavy machinery traffic. In 3 months, 2 of the 12 cameras had the lens covered in concrete dust that no algorithm cleaned. Solution: weekly cleaning kit + dome with integrated cleaner (wiper blade). Additional $800 USD not budgeted.
2. Monthly model retraining The site changes daily (new structure, new scaffolding, new work positions). The model fine-tuned with February data lost accuracy by May when the site changed from foundation to structure phase. Solution: automated monthly retraining pipeline with 200 new annotated images (4 person-hours/month). Additional $200 USD/month.
3. Privacy and staff resistance Workers initially rejected the system (perception of “Big Brother”). The local union raised objection. Solution: informational meeting + visible signs of “video-surveilled zone with PPE detection for your safety” + opt-out option for administrative staff (not operational). The system operates only in work zones, not in dining areas or bathrooms.
Lessons learned
- 87% F1 on critical class is sufficient for production, but requires monthly tuning. It is not fire-and-forget.
- The real ROI is not in the avoided fine, it is in contract retention. Construction companies that lose a contract for safety non-compliance lose $1-10M MXN; the system pays for itself with just one retained contract.
- Staff resistance is real and must be managed. An informational meeting + clear signage converts the system from “oppressive surveillance” to “preventive safety”.
- Edge AI > cloud AI for this case. <500ms latency critical for real-time alerts; cloud with 1-2s round-trip does not work.
Executive summary
For a construction site with 80-150 workers, 3-12 months duration, CV system for PPE has:
- CAPEX: $25,000-35,000 USD
- Annual OPEX: $4,000-5,000 USD
- Payback: 6-26 months depending on whether contract retention is counted or not
- Realistic accuracy: 85-90% F1 on critical class (no PPE)
- Dominant ROI if site has contract with safety compliance clause
P.D. Does your construction company have a contractual safety compliance clause? If yes, the CV system pays for itself with just one retained contract. If no, the ROI is 24-30 months on avoided injuries alone.
Frequently asked questions
How much does it cost to deploy EPP detection with computer vision on a 30,000 m² construction site? Typical CAPEX: $35,000-$60,000 USD. Components: 8-12 IP cameras (existing, no new hardware) + VMS licenses with EPP analytics ($800-1,500 USD/camera/year) + PoE switch + edge inference server ($4K-8K). Annual OPEX: 15-20% of CAPEX. Payback: 6-12 months via reduction of incidents that stop work (each unscheduled stop costs $15K-50K USD/day on projects >$5M MXN).
Does it work with non-standard helmet colors? Yes, with caveats. Pre-trained models (YOLOv8, YOLOv9) detect ‘generic helmets’ with 92-95% accuracy in daylight. For specific color helmets (green for visitors, red for supervisors), accuracy drops to 75-85% without re-training. The solution is fine-tuning with 200-500 images of your site: 1-2 weeks of work + $3K-5K USD from a data engineer. After fine-tuning, accuracy rises to 95-98%.
What about reflective vests? Same pattern. Orange/yellow vests are easiest (95%+ accuracy). Vests with large logos or reflective stripes can confuse the model because the stripe reflects the camera’s IR light and creates a ‘white’ that the model reads as ‘no-vest’. Solution: cameras with good IR reflection reduction, or switch to cameras with visible light + analytics that operate in the visible spectrum, not IR.
Do I need new cameras? Probably not. 80% of modern construction sites already have IP cameras for surveillance. EPP analytics runs as software on existing footage. If your cameras are 2020+ and have at least 1080p resolution, they work. If older or 720p, accuracy drops to 60-70% — not reliable for real-time alerts. Extra CAPEX to upgrade: $200-500 USD/camera × 12 = $2.4K-6K USD.
How many false positives does it generate? Without fine-tuning: 15-30% of detections are false positives (shadows, branches, birds). Operationally unsustainable — the operator learns to ignore all alerts. With fine-tuning: 2-5% false positives. Operationally viable. The trick is not the technology, it’s the fine-tuning with data from your specific site. Buying the most expensive camera on the market without fine-tuning = the same problem as a cheap camera with fine-tuning.
What about worker privacy? Critical. LFPDPPP in Mexico requires express privacy notice for ‘biometric data processing’. EPP detection doesn’t use biometrics strictly, but does use facial image for tracking if the camera is above the entrance. Solution: (1) cameras in work zones NOT in break zones, (2) visible privacy notice in each zone, (3) footage retained maximum 30 days, (4) worker’s right to request suppression of their image (uncommon but exists). Without these 4, you’re in legal gray area.
Sources
[1] OSHA — Personal Protective Equipment (PPE) standards 1910.135-1910.140 — https://www.osha.gov/personal-protective-equipment [2] BLS — Workplace injuries and fatalities construction sector 2023 — https://www.bls.gov/iif/ [3] Ultralytics — YOLOv8 architecture documentation — https://docs.ultralytics.com/models/yolov8/ [4] NVIDIA — Edge AI inference on Jetson Orin for industrial CV — https://developer.nvidia.com/embedded/jetson-orin [5] Roboflow — Construction PPE detection dataset and benchmark — https://universe.roboflow.com/ [6] STO Building Group — AI PPE compliance case study — https://www.stobuildinggroup.com/ [7] CITF Hong Kong — Smart Site PPE monitoring requirements — https://www.citf.cic.hk/