The Bureau of Labor Statistics has published the Census of Fatal Occupational Injuries every year since 1992. If you plot construction's fatal injury rate per 100,000 full-time equivalent workers from 2010 to 2024, you get a line that goes from about 9.5 to about 8.5. That's an 11% decline over 14 years. Manufacturing's rate dropped 31% over the same period. Warehousing dropped 28%. Construction barely moved.
The typical response to this data, when you present it to a safety officer at a large GC, is: "We've been working on this for years." That's true. The question is why effort hasn't translated into improvement at the same pace as adjacent industries, and whether the tools deployed explain the gap.
The Fatal Four haven't moved either
OSHA's "Fatal Four" — falls, struck-by objects, caught-in/between events, and electrocution — accounted for 58.6% of construction worker deaths in 2023. The exact same four causes, in roughly the same proportion, dominated fatality data in 2010. The specific mechanisms killing construction workers have not changed meaningfully in 15 years despite substantial regulatory activity, widespread safety training programs, and increasing toolbox talk frequency across the industry.
This suggests the problem is not lack of awareness. Workers know falls kill people. Site safety managers know crane swing radii are lethal. The gap is the interval between a hazardous condition forming and a supervisor intervening. That interval, on a typical commercial jobsite, is measured in hours — the time between scheduled walk-throughs. The hazard often resolves itself one way or another before the next inspection.
How manufacturing improved faster
Manufacturing's faster improvement trajectory is partly explained by fixed-station work. A press operator works at the same machine every day; that machine can be instrumented with sensors, guarded with physical barriers, and monitored continuously. The observation problem is solved by the nature of the work.
Construction is the opposite. The work environment changes daily. A form carpenter working at a fixed location on Monday may be working at a different elevation on a different part of the site by Thursday. The crane that was stationary yesterday is repositioned overnight, invalidating the exclusion zone that was drawn around its previous swing radius. Physical barriers are removed and reinstalled constantly. Any monitoring system that assumes static work positions will fail within a week on an active jobsite.
That dynamic nature is precisely why construction lagged behind manufacturing in adopting sensor-based safety monitoring. The technology that works in a factory — fixed-position sensors, perimeter guards, machine-linked interlocks — doesn't map to a construction environment. The industry needed a different approach, and computer vision with mobile wearable sensors is the first technology that actually fits the problem shape.
The measurement problem underneath the statistics
The 11% improvement figure is also partially a measurement artifact. OSHA 300 log accuracy is a known problem in construction. Recordable incidents are systematically under-reported, particularly among subcontractor workforces on large projects where the subcontractor's own TRIR affects future bid eligibility. When a subcontractor's foreman has strong incentives not to classify an injury as recordable, some incidents get reclassified as first-aid-only events that don't appear in the OSHA 300 log.
Continuous camera coverage changes this dynamic. When a safety detection event is timestamped and frame-referenced before anyone makes a reporting decision, the pressure to reclassify is reduced by the existence of objective evidence. Our pilot sites reported a 12% increase in first-year recordable incident logging after deployment — not because incidents increased, but because the baseline reporting was previously incomplete. That's a real data quality improvement even if it temporarily makes TRIR look worse.
The near-miss data problem
Leading indicators — near-miss frequency, PPE compliance rate, zone breach count — are more predictive of future fatal injuries than lagging indicators like TRIR. Heinrich's triangle (updated in later research by Frank Bird) suggested roughly 600 near-misses for every major injury. More recent construction-specific research suggests ratios between 300:1 and 400:1 for fatal events. Either way, the math tells you that if you want to move fatality rates, you need to intervene much earlier in the incident pyramid.
The problem is that near-miss data capture in construction is essentially broken. Voluntary near-miss reporting requires workers to self-report events that may reflect poorly on their own behavior, report to supervisors who may penalize them for the near-miss, and remember to file a report after a long shift. Our pilot sites, which ran continuous detection, captured 8-12x more candidate near-miss events per month than they had under manual reporting — as discussed in our piece on OSHA 300 log automation. That's the data infrastructure the industry needs before it can meaningfully improve leading indicator measurement.
Fall protection: the persistent gap
Falls account for roughly 36% of construction fatalities annually. OSHA's fall protection standard (29 CFR 1926.502) is among the most cited violations each year, alongside fall hazard training requirements. The regulation is clear, the violations are well-documented, and falls still kill more construction workers than any other mechanism.
The enforcement model is part of the problem. OSHA inspection rates for construction sites are low — far below what would be needed to create credible deterrence through spot checks. The effective compliance driver is not fear of inspection but the cost of incidents: workers' compensation claims, project delays, and the direct costs associated with a recordable fatality. Those costs are real but diffuse and delayed. They don't operate in the 30-second window when a worker decides to cross a floor opening without tying off.
Computer vision detection of fall protection non-compliance — workers near unguarded edges without lanyards, scaffolding access without proper harness attachment — operates in that 30-second window. The alert goes to the site supervisor before the worker has taken more than a few steps in the hazardous configuration. That's the mechanism for moving fall rates: not more regulation, but shorter observation cycles.
What's actually different now
Three converging developments make 2025 meaningfully different from 2015. First, computer vision model quality — specifically YOLOv8 and its successors — has reached the point where PPE detection accuracy on real jobsite footage exceeds 95% under daylight conditions. That wasn't true five years ago. The models that existed in 2018-2020 failed badly on hard hats with non-standard colors, partial occlusion, and low-light conditions.
Second, edge computing hardware has become affordable enough to put meaningful inference capability on-site rather than routing through cloud infrastructure with unacceptable latency. A modern edge server running 48 RTSP streams costs under $12,000 — significant, but within capital budgets for a commercial project.
Third, BLE mesh sensor technology has matured to the point where per-worker wearables cost less than $80 per unit and achieve 500ms update rates for position and biometric data. The combination of overhead camera coverage and per-worker BLE sensors provides two independent detection mechanisms, which matters for high-reliability alerting in environments where cameras have periodic occlusion.
None of this makes the problem easy. But the observation infrastructure now exists to actually measure construction safety at the granularity needed to improve it. That's what's different.
A realistic target
Manufacturing improved fatal injury rates by 31% over 14 years. That's a reasonable medium-term target for construction if the industry adopts continuous monitoring at scale. Getting there requires both technology deployment and a shift in how safety performance is measured: fewer lagging indicators, more leading indicators, and data infrastructure that captures near-misses before they become recordable events. The tools are available. The remaining constraint is adoption speed — which is ultimately a business decision by the GCs and owners who control site budgets.