Pedestrian Detection Systems in Construction: An EHS Field Guide
Why on-machine AI belongs on the iron that backs and swings through your crew, and how to prove one unit before you scale.
Introduction
A laborer steps behind a reversing excavator to set a grade pin. The operator checked the mirrors. The spotter is pulling another machine. The path the laborer is standing on didn't exist last Tuesday.
Nobody was reckless. The geometry was.
That is the construction version of a problem every safety program already knows and still fails to measure. Training is real. Experience is real. "Look twice" is real. None of it repeals a house swing, a spoil pile, or a visitor who was not on yesterday's roster. After a serious hit, someone always asks how many times this almost happened first. Most jobs cannot answer, because the close call never made the log.
Falls still kill more construction workers than any other single event. Getting hit by equipment is the part a camera on the machine can actually shrink. Especially inside the fence, where the iron is already on the job.
So let me be plain about what this article means by the term. Pedestrian detection systems in construction are vehicle-mounted AI vision. Cameras on the machine classify a person in a set zone and warn the operator in time to stop. The system sits on top of the internal traffic plan, spotters, and exclusion zones. It does not replace them. Get that framing wrong and the rest of the money is wasted.
Here is the promise of the next few thousand words. Why jobsites break warehouse thinking. How on-machine vision works and where it fails. How to prove one unit in mud and glare instead of on a clean pad. And how to keep a multi-employer crew from treating the display like a spy cam.
Two Different Ways People Get Hit on a Construction Site
Focus Four, Said Plainly
OSHA's Focus Four still drive most construction deaths: falls, getting hit, caught-in or between, and electrocution. Detection does not fix scaffolds. It does not touch trench collapse. It addresses one thing. People versus moving iron. Keep the scope honest and the tool stays credible.
Highway Traffic vs. Iron Already on the Job
Buyers mix up two different hits. Outside the workspace, a public vehicle comes through the cones. That is a temporary traffic control problem. Inside the workspace, a dump truck, an excavator, a telehandler, or a loader is already on the job. That is an internal traffic problem.
Vehicle-mounted detection is an inside-the-fence control. Pair it with an internal traffic control plan. Do not sell it as a substitute for cones, barriers, or flaggers.
NIOSH makes the split concrete. In a review of Fatality Assessment and Control Evaluation reports covering 78 roadway work-zone deaths across 75 reports, 41 incidents, about 55 percent, involved construction vehicles and equipment operating inside the work zone. Backing over people. Running over people. Caught in or between equipment. The danger was already on the payroll.
The Numbers Nobody Wants on the Board
The current counts are ugly enough without dressing them up.
Per the BLS Census of Fatal Occupational Injuries, construction and extraction occupations recorded 1,032 fatal injuries in 2024, and falls among those workers came to 370 and still led. Across all industries, pedestrian incidents involving motorized land vehicles rose sharply, up 19 percent to 369 in 2024 from 310 the year before. People on foot are getting hit more, not less.
CPWR's June 2026 Data Bulletin is the one civil buyers should read. In Heavy and Civil Engineering, transportation incidents were the single largest fatal event at 47.3 percent (88 deaths), and struck-by was the leading Focus Four fatal injury of the year at 28.5 percent, or 53 deaths. The hazard is the moving machine.
The older backing pattern has not gone away either. In a BLS review of the 962 fatal injuries at road construction sites from 2003 to 2010, 443 workers were struck by a vehicle or mobile equipment. One hundred forty-three of those were backing incidents, and dump trucks accounted for 84 of them. This is not new. It is just still happening.
"The recorded incident is the last frame. The first hundred frames never made the log."
That is why a clean lagging rate can sit right next to a job that is one step from a funeral. A low number on the board is not the same as low exposure frequency on the ground.
Why a Warehouse Spec Fails on a Jobsite
A system designed for a lit aisle map will lie to you on a cut. Here is why the transplant fails.
The Layout Won't Exist Tomorrow
Layouts change daily. Exclusion paint from Monday is a travel path on Wednesday. Grade shifts. Stockpiles move. Temporary roads appear overnight. Anything that needs ceiling infrastructure or a stable floor plan is the wrong architecture for dirt.
"The layout that was safe on Monday is a travel lane on Wednesday."
The Roster Changes at the Gate
General contractor, dirt crew, concrete, electrical, plumbing, HVAC, inspectors, delivery drivers, owner's reps. The roster changes at the gate every morning. Protection that depends on issuing hardware to every person who walks the site dies at orientation. The person most at risk is often the one who wasn't on site yesterday. If protection depends on handing out a tag at the gate, it fails by second shift.
No Building, No Network
Jobsites lack fixed power and reliable Wi-Fi. Detection has to run on the machine, powered by the machine, with no cloud hop inside the safety loop. If the radio dies, the warning still has to work. A dead radio cannot be part of the safety loop.
Operators Rotate Iron and Blind Spots Scale
The same person may run a skid steer at 7 a.m. and a telehandler at 2 p.m. The interface has to be obvious in one sitting. Training written for one assigned lift will fail. And blind spots scale with the machine. A warehouse counterweight is not an excavator house plus boom plus spoil pile. Large equipment needs three or four cameras, not a single rear puck.
Before you write a spec, answer the jobsite questions, not the brochure questions. What is the light, dust, rain, and glare on this cut? Do you need to know whether it is a person or a Jersey barrier? Who walks this site who will never be enrolled in anything, ever? What is the real all-in cost once you add mount, training, and lens cleaning to the sticker? And can the kit leave with the rental when the rental leaves? Answer those honestly, and half the vendors on your list disqualify themselves. If you want a second set of eyes on that spec, that is a conversation worth having before you buy anything.
What Pedestrian Detection Systems in Construction Actually Do
Cameras on the house, counterweight, and flanks watch the hole the operator cannot see. An onboard model reads body shape, posture, and movement. When a person enters a set zone, the cab gets a light and a tone. Event clips can stay on the machine and upload later. The system is supposed to scream for a person, not for a pallet or a barrier.
That is the whole point of pedestrian detection systems in construction. They put a set of eyes on the parts of the machine no operator can watch at once, and they turn a close call into something you can actually see on video.
How On-Machine Vision Works
Typical hardware from systems documented in industrial use runs wide-angle lenses around 90 to 120 degrees, useful detection often in the 15 to 30 foot band, and planning accuracy above 95 percent in good light. Expect that number to fall when the light falls. Treat every figure here as a planning range to prove on your iron, not a guarantee.
Why the Math Happens on the Machine
Do the math on the job, not in the cloud. At a modest 5 mph, a machine covers about 7.3 feet every second. Every extra delay is distance with no warning. On-device systems typically alert in about 50 to 100 milliseconds. Video stays local. Events can wait for a signal.
If Wi-Fi dies in a basement, on a steel floor, or in a highway median, detection keeps running. If the onboard box dies, the operator should know immediately. Silence should not be ambiguous.
Zones the Operator Can Live With
Keep the logic simple. Inner zone: stop now, often about 5 to 12 feet. Middle zone: slow and look, often about 12 to 20 feet. Outer band: tracking only, no tone.
Start wide, around 20 to 25 feet, so you miss fewer people. Then tighten after walk-tests so the operator does not drown in noise. Front zones are often larger than rear, because travel speed is higher going forward, and some systems grow the zone as speed rises. The point is not a pretty default. The point is a tone the operator still believes in week six.
How Many Cameras
Camera count follows the size of the hole. A skid steer or compact track loader usually needs two, rear and side, because the risk sits close and to the turning side. A telehandler wants three or four for the long load and yard travel. An excavator earns four, because the house swings and a crew works the trench toe. Wheel loaders and dumpers that reverse on haul roads run three or four. Cranes and large iron need four.
Put simply: one camera is rear-only. Two, front and rear, cover roughly 70 to 80 percent of the hole. Four, front, rear, left, and right, are the honest answer on big machines. A single rear puck is a warehouse answer to a construction question.
Say the Limits First
Vision needs light and a clean lens. Performance drops below about 20 to 30 footcandles. Accuracy above 95 percent in good light can drop to 70 percent or lower in the dark. Dust that cuts visibility under 20 feet will beat a camera. Rain, snow, and low sun off steel will beat it some days. A laborer behind a spoil pile is invisible until they step out. A dirty lens is a blind system.
Weatherized housings rated IP67 or better, vibration resistance, and a wide temperature range are table stakes at any price, not upgrades. Cost, though, depends entirely on architecture, and the spread is enormous. Older multi-camera enterprise systems ran about $8,000 to $15,000 per machine installed, plus $200 to $400 a year in support.
Newer self-contained vision kits have collapsed that math. A current single unit commonly lands in the low four figures, roughly $1,500 to $2,000, with no recurring support fee. Price it for the technology you are actually buying, not the last generation's brochure.
If the lens is mud, improve lighting, tighten the exclusion, put a dedicated spotter on that task, and re-test. Do not pretend the camera can see through a pile.
Match the System to the Phase, Not the Company Logo
The right spec changes as the job changes. What we're seeing across sites is that the buyers who freeze one spec for a whole build end up with too little coverage early and wasted cameras late.
Site prep, earthwork, and civil are the highest payoff. Excavators, loaders, dumpers, graders, and crews on foot at the cut. Start with the machines that have the worst holes and the most people walking through them. Foundation and structure bring concrete pumps, telehandlers, and cranes, and congestion spikes. That is three or four cameras on the large iron. Interior and MEP tighten the corridors and multiply the trades and visitors. Cameras still matter on material handlers, but layout control matters more than extra cameras there.
Hyperscale, data-center, and heavy industrial projects concentrate hundreds of workers, dozens of employers, and, at peak, 500-plus people and 50-plus vehicles. Owners write "zero incident" into the spec. The outdoor phases are where self-contained cameras on excavators, cranes, and concrete trucks earn their keep, because no site network can be leaned on. Phase the detection plan the way you phase the work.
Highway and road work is its own case. Pair detection with an internal traffic control plan that isolates people from equipment, reduces the need to back up, limits access points, and marks pedestrian-free zones. Detection covers the machine inside the cones. It does not stop the motorist outside them.
Validate One Machine in the Worst Dirt, Not the Cleanest Pad
This is the spine of the whole thing. Everything above is background. This is what decides whether you buy safety or shelfware.
Why Demos Lie
A parking-lot demo is theater. Level ground. No dust. No low sun. A safety intern walking the zone on cue. That proves the box can see a person. It does not prove it will work on your job.
What the demo hides is everything that matters. Dawn and dusk. Night work. Low sun off steel. Dust after a cut. Rain on the lens. Three trades and a delivery at once. Vibration on rental iron. And an operator who is already tired of gadgets. Treat a demo as a screen, not as validation. Demo it in the cut, not in the job trailer.
Why "Free 90-Day Pilots" Lie
The box can be free and still cost you. Coordination, training, and babysitting eat 40 to 80 hours of your people. The vendor mounts it, tunes it, and reports the pretty numbers. The best operator gets the cleanest pad in the best weather. After 90 days, the project wants to finish what it started. And vendor-level service tends to disappear the day you buy twenty more.
A pilot only helps if you get raw logs, your mechanics do the work, you test the worst weather, and you wrote pass and fail before day one. Most vendor pilots fail that test.
Buy One Unit
Purchase one complete kit at full price. Your people mount it, or you stand over the installer. Your operators run it. Your mechanics clean it. Run 30 to 90 days across shifts and weather. A modern self-contained kit runs roughly $1,500 to $2,000, so proving one is a rounding error next to a fleet nobody uses. Legacy enterprise units cost several times that, which only makes testing one before you commit more important. This single-machine validation is the same discipline we use before anyone signs off on a fleet.
Put it on the machine that actually hurts people. Usually the excavator or dumper that reverses through a crew. Put a respected operator in the seat. If that person wants it, the rest of the fleet will listen. If that person wants it dead, believe them.
What to Collect and the Pass or Fail Bar
Count alerts per shift. Review video and the operator log to separate true alerts from false ones. Track uptime. Track lens-clean and repair time. Once a week, sit for 15 to 20 minutes and ask the same questions. Did an alert help you this week? Any false alerts that burned trust? What would you change? Score trust from 1 to 10 and watch the trend. Also log the close calls the system caught, the ones it missed, and any hit with the system on.
Write pass and fail before day one. In documented industrial practice the bar looks like this. Detection in the set zone above 95 percent by video sample. False alerts under 10 percent while the system settles, under 5 percent once it is stable. Most operators who touch the machine would want it on their regular iron. Maintenance under two hours a month per machine. Jobsite dirt may blow that last number, so treat it as a design constraint, not a wish.
If it passes, instrument 5 to 10 high-risk units. If a critical item fails, stop or change the spec. Do not learn your way into a fleet of ignored alarms.
A Seven-Step Playbook
Map the job. Light, dust, weather, machine classes, peak congestion, and the worst close-call spots. Measure footcandles in the dark corners.
Call three customers on similar dirt. Ask false-alert rate, lens-clean load, and whether they would buy again.
Demo on this ground. Person from a blind angle. Person at the trench toe during a house swing. Two people in the zone. Worst glare. Worst dust.
Buy one unit. Full data access. Support during the test. Volume price only if the test passes.
Mount it yourselves, or watch every bolt. Document every fight with the machine. That fight predicts fleet pain.
Run the clock. Learning weeks will be noisy. Stable weeks are the real data. Add a stress week: night, rain, peak congestion.
Write the engineering note. Metrics versus the pre-set bar. Go, no-go, or try another box. The sentence leadership needs is simple. We have X weeks on the worst machine in our weather.
Stress what will actually fail you. Darkest night work. Dustiest cut. Low sun off wet ground. Heat and cold the job sees. A person crouched or kneeling. A person in a dirty vest stepping out from behind a pile. Two operators over two weeks, because this crew rotates iron.
Installation, the Business Case, and the Culture That Keeps It Alive
Installing on Rental Iron
Power the system from the machine. Do not trench for cable, and do not open an IT ticket. Spec magnetic mounts and quick-disconnect so the kit follows the rental off the job. Standardize zone logic by machine class, because an excavator is not a telehandler.
Phase 1 is not the fleet. Phase 1 is the iron that reverses or swings through people.
You have three install paths. The shop does it, which is cheapest, slowest, and leaves the most knowledge on the job. A hired crew does it, which is fastest and leaves the least behind. A hired crew trains the shop and leaves, which is the middle path. A handful of machines favors the shop. A 20 to 50 mixed fleet favors train-the-shop. A 50-plus or multi-job rollout favors a hired crew if the calendar is on fire.
Aim cameras at the actual hole, not the brochure photo. Automotive wire, loom, and a fuse near the source. Power off the switched ignition. Walk-test every mount so a person entering from any angle is seen and parked objects stay quiet. Recalibrate when the layout or the season changes. After the first unit, leave a job book: a photo mount guide by machine class, baseline zones, what you changed from factory defaults and why, a 15-minute operator card, and a lens-clean schedule.
The GC and Owner Memo
You do not buy this because a spreadsheet says a life is cheaper than a kit. You buy it because people who show up for work are not supposed to get run over. The memo exists so the budget can catch up with that fact.
Planning numbers, labeled as planning ranges, and they swing hard by architecture. A modern self-contained kit runs roughly $1,500 to $2,000 per machine with no recurring support fee, which puts a first package of 5 to 10 high-risk units somewhere around $7,500 to $20,000. Older multi-camera enterprise systems still run $8,000 to $15,000 per machine installed, plus $200 to $400 a year, and push a comparable package well into six figures. Either way, phase it. Prove the class that works, then spread the rest across 12 to 24 months. At modern-kit pricing the capital argument mostly disappears, but financing exists if the year cannot swallow the whole list.
Offsets you can name without pricing a femur. Experience modification and carrier posture. Some carriers talk 5 to 15 percent credits on qualified loss-control technology. Staying insurable matters more than the discount. After a serious event, investigators look at whether you used available engineering controls on a known hazard. Owner prequal and "zero incident" language on industrial work. Video instead of three conflicting statements. Some operations report less hesitation at blind corners, but treat flow talk as a maybe. It is not the reason.
Keep the owner memo to five short paragraphs. This job runs named machine classes through people on foot, and training and cones have not closed the hole. The control is vehicle-mounted AI on those classes, on top of the traffic plan. Phase 1 is one worst machine for 90 days, then 5 to 10 units, at a stated dollar range. What the job gets is close calls you can see, video when something happens, and kits that leave with the rental. The ask is approval of the first package, with a full fleet only if the test passes. Do not price a femur. Fund Phase 1. If it helps, sit down with your team and build that Phase 1 spec and dollar range against your actual machine list.
Then measure leading indicators, not injuries avoided. Alerts by zone and shift. Close calls caught. Close calls missed. Operator trust. Lens-clean load. That is operational intelligence you can act on before someone gets hurt, not a lagging report you read after.
Why the Box Dies at the Gate
Good technology, bad rollout, and operators kill it in months. Not because they hate safety. Because they think it is a judge they did not help pick.
Say "this watches out for you." Do not say "this watches you." Tell trades, before the first mount, that isolated alerts are not a write-up tool, that persistent patterns can be coaching, and that scoring productivity off the display is how you get tape on the lens. If the project manager never asks how the system is working, the crew decides it is optional. If a superintendent mocks the beeper, the fuse is pulled by week two.
Train in three blocks. Forty-five to 60 minutes on why: people, not features; what the tones mean; and what the system cannot see. Show one real close-call clip from this job if you have it, and admit false alerts exist. Then 30 to 45 minutes in the seat, with a trainer walking the zone so the operator hears the real tone while the machine is running. Then two or three shifts with a coach nearby and a check-in at break. Expect 10 to 20 percent early resistance and fix it one-on-one, with a respected operator as the advocate.
Week one, every beep is new. Month six, the brain filters it. Too many false alerts train people to ignore the real one. Review the log every quarter. Re-aim with the season. Make a dirty lens a pre-start fail, same as a dead backup alarm. And keep the human controls. Spotters. High-visibility clothing. Eye contact before crossing. Nobody walking the swing radius. New-sub orientation should treat the display like the vest, part of the site rules.
A 90-Day Plan
Days 1 to 14. List iron by back- and swing-exposure. Walk the three worst interactions with a superintendent and an operator. Write pass and fail. Measure light in the dark corners.
Days 15 to 30. Call three references on similar dirt. Demo in the cut, not in the job trailer. Buy one kit. Mount it with the operator in the seat.
Days 31 to 75. Run all shifts. Clean lenses on a schedule. Interview two operators a week. Log alerts, misses, and downtime.
Days 76 to 90. Score against the bar. Scale the class that worked. Change the traffic plan where vision cannot see. Do not put a system on the whole fleet if the first operator wants it dead.
From the Author
I have spent about thirty years around machine control, positioning, and pedestrian safety, and the pattern rarely changes. The technology is usually good enough. The rollout is where it lives or dies. The jobs that succeed are the ones that picked the worst machine, proved it in their own dirt, and let a respected operator carry the message to the crew. The jobs that fail bought a fleet on the strength of a clean demo and then wondered why the alerts got ignored by spring.
The other lesson is about measurement. Most organizations discover, only after a serious event, that they were blind to how often the near miss was happening. On-machine vision does not just warn in the moment. It gives you a record of the exposure you could never see before, which is the thing that lets you fix the traffic plan instead of the paperwork. I write about this in more depth at johnbuttery.com.
Why This Matters Now
In heavy and civil work, transportation incidents are the single largest fatal event, and struck-by is the leading Focus Four fatal injury. The dump-truck backing pattern that killed ground crews twenty years ago has not gone away. Cones and training are necessary and they are not closing the hole. What has changed is that a self-contained camera can now run on the machine, in the dirt, with no network, and hand you leading indicators instead of a post-incident report.
The shift for EHS and operations is from counting incidents after the fact to measuring exposure frequency as it happens. That is the real value here. Not compliance for its own sake, but operational intelligence you can act on. If you want a practical primer on that mindset, the note on proving safety technology before you scale covers the same discipline applied across mixed fleets.
FAQ
What is a pedestrian detection system on construction equipment?
Vehicle-mounted cameras plus on-device AI that warn the operator when a person enters a set zone around the machine.
Does OSHA require AI cameras on excavators?
No specific mandate. 29 CFR 1926 still governs vehicles and equipment. Detection is an extra engineering control on a known hazard.
Does this replace spotters and an internal traffic plan?
No. NIOSH treats separation and an internal traffic control plan as the primary inside-the-fence control. Detection is extra eyes on the machine.
Which machines first?
The ones that reverse or swing through people: excavators, dumpers, large telehandlers, cranes.
How many cameras does an excavator need?
Plan on four. Compact machines can start at two.
Will it work in dust, rain, and night work?
Only with line of sight and a clean lens. Test in your weather. If the lens is mud, the system is blind.
How long should a jobsite trial last?
One machine, 30 to 90 days, worst conditions, your operators, your mechanics.
What should a first unit cost?
A modern self-contained kit runs roughly $1,500 to $2,000 with no recurring support fee. Older multi-camera enterprise systems run $8,000 to $15,000 installed.
Can the kit move with rental equipment?
Yes, if you spec quick-mount hardware and keep the kit on the asset log.
Conclusion
The job will change tomorrow. The visitor will not be on yesterday's roster. The operator can look in only one direction at a time. Training does not repeal any of that. What you can do is put eyes on the machine that actually backs through people, prove one unit in the dirt, and scale only what survived the test.
"Put the first kit on the machine that actually reverses through a crew, and let the operator tell you the truth."
How Riodatos Helps
The whole argument of this article is validation before scale, and that is the work we do with safety teams. We help you pick the one machine that matters, run a single-machine evaluation in live jobsite conditions, and hold the results against a pass-fail bar you set before day one. No fleet commitment until the worst machine, in your weather, earns it.
If that approach fits how you buy, start with live jobsite validation, reach us through contact, or book a 30-minute call at calendly.com/john-buttery-riodatos/30min. You can also browse our services to see how evaluation, install, and support fit together, and read the construction companion note on collision avoidance around construction vehicles.
About Riodatos
Riodatos is a U.S.-based industrial safety technology company headquartered in Arizona, with domestic inventory and hands-on support for vehicle-mounted pedestrian and proximity detection across the Americas. We supply, configure, install, and support systems matched to site-specific equipment, traffic, and risk, whether the work is a warehouse, a factory floor, an active construction site, or a mixed logistics fleet.
Our focus is measurable live performance, operator adoption, and deployments that scale across mixed fleets and multiple sites rather than mismatched technology or overseas delays. Direct pricing, fast U.S. shipping, certified installation, and English and Spanish support let safety teams put protection first and prove it before they scale. Learn more at riodatos.com or read the blog for more field notes.
Sources
BLS, Census of Fatal Occupational Injuries, 2024 (1,032 construction and extraction fatalities; 370 falls; 369 pedestrian vehicle incidents, up 19 percent from 310): https://www.bls.gov/news.release/cfoi.nr0.htm
CPWR, June 2026 Data Bulletin, Construction Injury Trends by Major Subsector (Heavy and Civil Engineering: transportation incidents 47.3 percent, n=88; struck-by leading Focus Four fatal injury, 28.5 percent, n=53): https://www.cpwr.com/wp-content/uploads/DataBulletin-June2026.pdf
NIOSH Science Bulletin, Using Internal Traffic Control Plans to Prevent Worker Injuries in Work Zones (75 FACE reports, 78 fatalities, 41 or 55 percent inside the work zone): https://www.cdc.gov/niosh/bulletin/2024/struck-by.html
BLS Monthly Labor Review, An Analysis of Fatal Occupational Injuries at Road Construction Sites, 2003 to 2010 (962 fatalities; 443 struck by vehicle or mobile equipment; 143 backing; 84 dump trucks): https://www.bls.gov/opub/mlr/2013/article/an-analysis-of-fatal-occupational-injuries-at-road-construction-sites-2003-2010.htm
NIOSH Workplace Solution 2014-125, Preventing Worker Injuries and Deaths from Backing Construction Vehicles and Equipment: https://www.cdc.gov/niosh/docs/wp-solutions/2014-125/pdfs/2014-125.pdf
OSHA, Focus Four and fall-prevention campaign: https://www.osha.gov/stop-falls
OSHA, Preventing Backovers resources: https://www.osha.gov/preventing-backovers
29 CFR 1926 Subpart O; 1926.602(a)(9)(ii), backup alarm as a baseline control, not a sufficient one
Framework paraphrased from John Buttery, Preventing Pedestrian Collisions (2026). Illustrative scenes in the book are fictitious. Dollar and accuracy figures are labeled as planning ranges to prove on your own equipment.
All statistics above were verified against the linked primary sources as of September 2026. Performance and cost figures are planning ranges, not guarantees.
