Computer Vision on Construction Job Sites: What's Actually Deployed

Computer vision construction tools are used mainly for progress tracking, where 360-degree cameras compare as-built conditions to BIM and the schedule, as with OpenSpace and Buildots. They also flag missing PPE and hazards, and document defects and quality. As of 2026, accuracy figures are vendor-reported and performance drops in poor lighting and crowded scenes.

Our take: capture is solved, interpretation is not. Walking a site with a helmet camera and getting a navigable record is established technology. The "AI" value depends on comparing what the camera sees to a plan and a schedule that are themselves accurate, and nobody has published an independent test of how well that works. For the broader market, see AI in construction in 2026.

What does computer vision do on a construction site?

Three use cases cover nearly everything deployed or studied.

Bluebeam's July 2026 survey found 28% of US AEC firms with AI in production, but it does not break out computer vision specifically, so we cannot say what share of those deployments involve cameras.

How does 360-degree progress tracking work?

The workflow is similar across vendors. A worker wears a camera on a hard hat and walks the site on a regular route. Software stitches the imagery, places it on the floor plan and, in AI-enabled products, identifies building elements and installation status.

OpenSpace, per trade-press coverage, uses helmet 360 capture and segmentation, then maps results to floor plans and trades. The company claims more than 4 billion square feet captured. That is a volume-of-use claim from the vendor, not an accuracy metric.

Buildots raised $45 million in May 2025 and compares hard-hat camera output to the plan and schedule. The difference in emphasis matters: one is described primarily as capture and mapping, the other as plan-versus-reality comparison. Even so, both claims are descriptions, and no independent benchmark comparing them was found.

Why does the plan matter more than the camera?

A camera can tell you a wall is framed. To say "this wall is on schedule," the software needs to know which wall it is, when it was supposed to be framed, and whether it is complete or only started. That requires a reliable model or plan set, a schedule with sensible activity IDs, and a mapping between them.

This is where the data problem reappears. If the BIM lags the design, or schedule activities do not correspond to physical elements, the comparison is noisy. Autodesk and FMI estimated bad data may have cost the global industry $1.85 trillion in 2020, and Bluebeam found firms with AI in production put integration with existing systems first (47%) among success factors. See the construction data problem.

The practical consequence: the best conditions for progress AI are projects with disciplined models and schedules, and those projects are already the better-run ones. Gains there may reflect the discipline, not only the tool. Related reading on the schedule side is in AI construction scheduling.

How accurate is computer vision for PPE and hazard detection?

The academic literature gives numbers, with warnings. Studies of PPE detectors report mean average precision (mAP) in the range of roughly 0.70 to 0.95 on curated datasets. Examples include 0.841 mAP@0.5 under variable lighting and 86.55% for YOLOv5x on a helmet dataset.

Do not compare these numbers across studies. Datasets differ in image quality, number of workers per frame, camera angle and annotation rules. A model that scores 0.95 on a clean helmet dataset may do much worse on your site at dusk. The cited research flags the same limitations: lighting, occlusion and worker clustering.

A second limitation is scope. Most detectors find helmets and vests only. They do not verify correct wear, such as a chin strap fastened or a harness properly connected. Detecting presence is not proof of compliance, a point explored in AI for construction safety.

Where does computer vision break down?

Recurring failure conditions across the sources:

• Lighting. Early morning, evening, glare and indoor-outdoor transitions reduce detection accuracy.
• Occlusion. Stacked materials, scaffolding, formwork and equipment hide the thing being measured.
• Crowding. Clustered workers merge into single detections or hide each other.
• Concealed work. Cameras see surfaces. In-wall rough-in, below-slab utilities and fireproofing behind finishes need other verification.
• Reference quality. Wrong or outdated plans produce confident, wrong comparisons.

Illustrative example: a tool reports 82% of a floor's drywall hung. The superintendent knows two bays are hung but not taped, and the north corridor was left open for a late inspection. The percentage is a useful prompt for a conversation, not a pay-application number. Who defines "complete" is an operational decision the software inherits from whoever configured it.

What about defect detection and quality documentation?

The brief for this topic often includes automated defect detection, so it is worth being precise. The sources we gathered support visual documentation as a use: repeated imagery gives a dated record of what a space looked like at a point in time, which helps with punch lists, warranty questions and disputes over who damaged what. We did not gather evidence on tools that automatically classify defects such as cracks, misaligned finishes or incorrect installation, and we make no accuracy claims about them.

The distinction is useful when evaluating a product pitch. "We record everything" is a claim about coverage and storage. "We find problems for you" is a claim about interpretation, and it needs a test. If a vendor says its system detects defects, ask for precision and recall on a labeled dataset from projects like yours, and ask how many false alarms a superintendent should expect per walk. A tool that raises 30 questionable flags per floor can consume the time it saves.

How should you evaluate a computer vision vendor?

Because independent benchmarks do not exist, the evaluation has to be yours. A structured pilot has five parts.

1. Define ground truth. Choose a floor or zone and have the project team record actual status by hand on the day of capture.
2. Capture on your normal cadence. A one-off demo walk by the vendor's best operator tells you little about weekly capture by a busy project engineer.
3. Score the output. Compare tool-reported status to ground truth at the element or activity level, and count false positives and false negatives separately.
4. Test the bad days. Include a dim interior, a crowded floor and an area with heavy temporary protection.
5. Time the whole workflow. Include capture, upload, processing delay and review, not only the headline analytics.

The cost side deserves the same rigor. Half of Bluebeam's respondents named implementation cost as their top barrier, so count camera hardware, licenses, the time to walk and the time to correct data, then measure value against that, as described in the baseline approach in measuring AI ROI in construction.

What does this mean for GCs, owners and designers?

GCs. Treat computer vision as documentation first, analytics second. A reliable capture record has value on its own for coordination, closeouts and dispute defense, regardless of how smart the interpretation is. Decide in advance who owns the walk route and cadence, since a tool is only as current as its last capture.

Owners and developers. Ask what the progress percentages are based on and who verified them. Use capture records to cross-check pay applications, but not as the sole basis for payment.

Designers. Repeated 360 capture can give design teams a visual record of conditions for RFIs and field-condition questions, saving site visits. It does not replace verification of conformance with design intent.

Each of these uses should be tested using a baseline-first pilot, covered in measuring AI ROI in construction.

What to watch through 2027

• Independent benchmarks. A third-party comparison of progress-tracking accuracy against surveyed ground truth.
• Vendor accuracy disclosure. Whether claims shift from square feet captured to element-level precision and recall.
• Beyond helmets and vests. Detectors that verify correct PPE use or detect fall-protection status.
• Lighting robustness. Research results on low-light and crowded-scene performance.
• Privacy and consent norms. Camera-based monitoring raises worker privacy questions; watch for owner and labor requirements.
• Adoption signal. Whether surveys such as Bluebeam's begin reporting computer vision as its own category, which would show how much of the 28% production share involves cameras and how much is document work.