Industries / Construction and property / Commercial and residential construction / ConTech
Safety
Safety technology is the equipment and software builders buy to warn, watch over and protect workers on site: tags, radar and cameras that sense people near machines, wearable sensors and exoskeletons, better helmets, and programs that read site photos and reports for hazards.
7
Milestones achieved, from 2006 to 2024.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2006
Radar tested on blind spots
NIOSH, the US federal institute for workplace safety research, spent two years testing a commercial radar warning system on large dump trucks at a surface mine and found that it reliably detected people, small vehicles and berms (earth banks at a road's edge) in the driver's blind spots, but often sounded for objects that posed no danger. In 2007 it tested a tag system for construction machines, in which a worker's tag senses a magnetic field from a reversing machine and both worker and operator get an alarm within about 15 feet (4.6 m); early tests at a road construction site found that it detected workers near reversing equipment.
LessonA warning that also sounds for harmless objects needs a way to check it, so NIOSH advised pairing sensors with cameras that let the driver see what set off the alarm.
Credit
Todd Ruff and L. D. Frederick, NIOSH
2016
A smart helmet that failed
DAQRI showed its Smart Helmet, an industrial helmet that laid work instructions and safety information over the wearer's view (augmented reality) and carried thermal sensors, during Intel's keynote at the CES trade show in January 2016. DAQRI raised $275 million in all, but TechCrunch reported that it struggled to win business customers, and in September 2019 it told customers it was ending its industrial wearables business and winding down.
LessonFunding and a stage demonstration did not make the product sell; TechCrunch's sources named the difficulty of training workers to use the hardware.
Credit
DAQRI (chief executive Brian Mullins), with Intel
2020
Tool makers sell exoskeletons
Hilti, a tool maker, launched the EXO-O1 with Ottobock in 2020: a frame worn like a backpack that uses springs and hydraulics, not electric motors, to hold up the arms during overhead work, which Hilti said relieves the shoulders by up to 47 percent, citing unnamed independent studies and Ottobock's own research. Outside studies have since found a mixed picture: a lab test of four arm- and back-support exoskeletons on 28 people found balance changes that suggest a higher risk of falls after slips and trips (2024), and in a study by CPWR, the research and training arm of the building trades unions, 60 percent of 22 workers said the devices eased discomfort while 79 percent of 24 industry interviewees worried about them snagging on objects (2025).
LessonA device that relieves one body part can add other risks, such as poorer balance and snagging, so tests need to look beyond the muscle it targets.
Credit
Hilti and Ottobock (product); S. Dooley, S. Kim, M. L. Madigan and M. A. Nussbaum (balance study); CPWR, with Mohamad Behjati Ashtiani, Maury Nussbaum and others (adoption study)
2022
AI reads site photos
Smartvid.io, founded in 2015, built an AI called Vinnie that was trained to spot people, hard hats, tools and equipment in site photos and to flag 100 kinds of hazard, and builders such as Suffolk Construction used it. The company renamed itself Newmetrix, and in December 2022 Oracle announced that it had completed buying it, to add the AI to Oracle's construction software.
Credit
Smartvid.io, later Newmetrix (founder Josh Kanner); Suffolk Construction (early user); Oracle
2023
Hong Kong mandates smart sites
Since February 2023 Hong Kong's Development Bureau has required public works contracts worth more than HK$30 million to adopt its Smart Site Safety System, which joins smart helmets that locate workers, AI cameras, and alarms for the danger zones of mobile machines and tower cranes to one central platform. From July 2024 the Buildings Department required the machine and crane alarms on private building works over HK$30 million, and from July 2025 on demolition, excavation, foundation and site formation works as well.
Credit
Development Bureau and Buildings Department, Hong Kong; Construction Industry Council (labelling scheme, from May 2024)
2023
OSHA switches to safety helmets
In December 2023 OSHA, the US workplace safety regulator, said its own staff would replace traditional hard hats with modern safety helmets, which have chin straps and protect the sides of the head as well as the top, because a strapless hard hat can fall off in a slip or trip and, by Bureau of Labor Statistics figures for 2020, about 20 percent of head injuries that cost workers days away from work came from slips, trips and falls. Its bulletin tells construction employers to consider Type II helmets (rated for blows to the top and sides) with chin straps, while leaving the choice to each employer's own hazard review.
Credit
Occupational Safety and Health Administration (OSHA), US Department of Labor
2024
Singapore requires site cameras
From 1 June 2024 Singapore's Ministry of Manpower has required every construction site with a contract value of S$5 million or more to install video surveillance where high-risk work happens, including work at height over 2 metres, excavation, lifting, and areas where excavators or forklifts are used. Footage must be kept for at least 30 days (180 days after a reportable incident), and the ministry says the cameras deter unsafe behaviour and help with training and investigations.
Credit
Ministry of Manpower, Singapore
Watch for
Hong Kong credits smart sites
Hong Kong's Secretary for Development said that since rules began requiring sites to adopt the Smart Site Safety System or its devices, the construction industry's overall accident rate fell about 20 percent from 2023 to 2025. She also said about 60 percent of new construction and maintenance projects over HK$30 million had adopted the system and received labels. The reply links the fall to the rules by timing only and gives no method for separating it from other causes.
Achieved whenan independent group, such as a university team using site-level accident records, compares sites with and without the system and finds fewer accidents where it is used.
Claimed by
Bernadette Linn, Secretary for Development, Hong Kong government, June 2026
Excavators that report near misses
Sumitomo Heavy Industries and NEC say they are building a system that uses camera footage and sensor data from hydraulic excavators to spot near misses (events that could have caused an accident but did not) and write the report automatically. A proof of concept in September 2025 produced reports of near-miss cases; joint development was due to begin in April 2026, with practical use as the goal in fiscal year 2027. They have published no accuracy figures.
Achieved whenthe system is in routine use by contractors outside the two companies' own trials, or an outside party publishes how many real near misses it caught and how many it missed.
Claimed by
Sumitomo Heavy Industries and NEC, March 2026
AI cameras around machines
Leica Geosystems says its Xsight360 uses up to six cameras for a 360-degree view to detect people or objects near a machine and alert the operator, with AI models from its partner Presien that were trained on 700,000 hours of site operations. The companies have published no figures for accuracy, missed people or false alarms.
Achieved whenan independent tester publishes detection and false-alarm rates for people standing and crouching around a machine fitted with the system, for example using the test methods in ISO 16001 or ISO 21815-1.
Claimed by
Leica Geosystems and Presien, June 2025
Open
01
Answer key for safety cameras
A public set of short video clips, filmed on a mock site around a parked machine, of consenting volunteers and mannequins. Some wear hard hats and high-visibility vests and some do not; some stand inside a marked exclusion zone (a no-go area around the machine) and some outside it. Each clip comes with a written answer key. Then run several AI systems over the clips and publish how many cases each got right, missed or flagged wrongly.
The briefCredit · unclaimed02
Near-miss report sorter
A public test set of written near-miss reports (short notes about something that nearly hurt someone) with an answer key, and a published tally of how well AI models sort them. Each report is labeled as an unsafe act, an unsafe condition or a good observation, the categories many sites use, and flagged if it could have caused a serious injury or death.
The briefCredit · unclaimed03
Parked machine alarm map
A measured map of where proximity warning systems (worn tags or AI cameras) do and do not sound around a parked excavator, using mannequins or tags on poles in place of people.
The briefCredit · unclaimed
Open 01 · the brief
Answer key for safety cameras
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of short video clips, filmed on a mock site around a parked machine, of consenting volunteers and mannequins. Some wear hard hats and high-visibility vests and some do not; some stand inside a marked exclusion zone (a no-go area around the machine) and some outside it. Each clip comes with a written answer key. Then run several AI systems over the clips and publish how many cases each got right, missed or flagged wrongly.
Why it is still open
Researchers wrote in July 2026 that publicly shareable construction-video benchmarks "remain scarce," especially for hazards that are dangerous to stage; their own set covers one hazard, a worker under a suspended load, in 55 synthetic clips. A March 2026 comparison of 10 safety video vendors by the analyst firm Verdantix is sold as a paid report, and its public page does not say how the vendors were tested. The open hard hat and vest set from Ultralytics, Construction-PPE, has 1,000 still images and no exclusion zones or video.
Built with what exists
Phone or site cameras; an open labeling tool such as CVAT or Label Studio; Ultralytics YOLO26, an open object detector, trained on the Construction-PPE set; a small vision-language model (an AI that reads images and text) fine-tuned on one GPU to label the clips (see the Note "One GPU to fine-tune"); trial accounts from commercial camera vendors that agree to be scored. The advance that makes this a small-team job is that open detectors and small vision-language models can now be trained on one graphics card.
First step
This month, mark a 5-metre zone with cones around a parked machine at a training yard or rental depot (engine off, key removed, nobody operating it), get written consent from five volunteers, and film 20 clips to test the answer key format. Filming people needs their written consent.
Done looks like
200 clips of 10 to 30 seconds, with an answer key that lists every person, whether each wears a hard hat and vest, and whether each is inside the zone; at least three systems scored; a published table of hits, misses and false alarms for each; clips and key released under an open licence with faces blurred.
Open 02 · the brief
Near-miss report sorter
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of written near-miss reports (short notes about something that nearly hurt someone) with an answer key, and a published tally of how well AI models sort them. Each report is labeled as an unsafe act, an unsafe condition or a good observation, the categories many sites use, and flagged if it could have caused a serious injury or death.
Why it is still open
A 2022 open-access study from IIT Madras says such reports "often require substantial manual classification efforts," and that limits on sharing near-miss data shaped its choice of site. It used private reports from one site in Kuwait; its best standard model scored an F1 of 0.79 (F1 is a 0-to-1 score that balances misses against false alarms) and struggled to tell unsafe acts from unsafe conditions. I found no public, labeled set of construction near-miss reports.
Built with what exists
Gemini Decision 1, a model that sorts messages, for example by whether the sender describes something dangerous (see the Note "A thousand questions, one pass"); a second large language model for comparison; a logistic regression baseline in scikit-learn, the kind of model that scored 0.79 in the 2022 study; Label Studio for the human labels; and, for spoken reports, ElevenLabs Scribe v3 speech-to-text (see the Note "Fewer word errors"). The advance is models that can label thousands of short messages in one pass at low cost.
First step
This month, have volunteers write 100 sample reports from their own experience, or get a contractor's written permission to use 100 anonymized report cards, and have two people label them separately to test the categories.
Done looks like
600 reports, each labeled by two experienced safety people working separately, with their rate of agreement published; F1 scores for each category for at least three models, set against the 0.79 baseline; the set released with all names, firms and sites removed.
Open 03 · the brief
Parked machine alarm map
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A measured map of where proximity warning systems (worn tags or AI cameras) do and do not sound around a parked excavator, using mannequins or tags on poles in place of people.
Why it is still open
INRS, the French institute for workplace safety research, wrote in September 2024 that the state of the art "does not at present allow a level of reliability to be guaranteed" for these detection devices, and asked that their missed detections and limits be described precisely, for people in any position, standing or crouching. The public field tests I found date from before today's systems: NIOSH's in the mid-2000s and the Construction Industry Institute's in 2010.
Built with what exists
A parked machine with the engine off and the key removed, or a full-size timber mock-up of one; two or more systems borrowed from dealers, such as a worn-tag system and a camera system like Leica Xsight360, powered from a separate 12-volt battery where the maker allows; adult-size mannequins in high-visibility clothing, posed standing and crouching; a tape measure and a chalk grid; phone video to record each alarm. The advance is AI that picks out people in the video from cameras mounted on the machine, now sold alongside tag systems; Leica Xsight360 is one example.
First step
This month, ask a local equipment dealer or operator training school for one hour with a parked machine and a demonstration unit, and draw the 72-point grid plan.
Done looks like
For each system, alarm or no alarm at 72 marked points (every metre out to 12 metres along six directions), three repeats each for a standing and a crouching mannequin, plus a count of alarms set off by 20 objects that are not people; published as a map and a table.