Construction Site Video Datasets: Equipment Activity, Progress and Crew Tasks

What construction video adds that progress photos cannot

Video is the right source when the target is motion over time: equipment cycles, idle time, crew sequences and progress rates. Still images answer "what is installed today"; a clip answers "how long did each bucket cycle take and why did the truck wait." If your model only classifies installed elements or PPE in single frames, our guide to construction progress photo datasets is the closer fit. The broader video data hub covers the checks that apply to every video category.

Three capture types dominate jobsite footage, and each produces a different dataset:

• Fixed jobsite cameras. Pole- or crane-mounted units that record time-lapse frames at set intervals or continuous streams. Strong for progress monitoring and site logistics; weak for fine-grained equipment state because machines are small in frame.
• Equipment-mounted cameras. Cab, boom or rear cameras on excavators, loaders and haul trucks. Strong for cycle recognition and operator assist; the field of view is constrained and vibration is common.
• Wearable and handheld cameras. Helmet or chest cameras on crew members. Useful for task recognition and procedural models, but they raise the heaviest consent questions because the wearer and nearby workers are recorded continuously.

Equipment activity labels: what the research shows

Equipment activity recognition is the best-studied construction video task, and most published datasets are narrow. Academic excavator and haul-truck datasets often cover a handful of sites, a few camera angles and a three-class scheme (digging, swinging, dumping). Some research sets also mix footage from public video platforms with site-recorded clips, which a commercial buyer cannot license. Read the data availability section of any paper before assuming its dataset can be used for a product.

For production models, specify the label set beyond the three-class baseline. Excavator states typically need idle, travel, dig, swing-loaded, dump and swing-empty, plus an "occluded" or "unknown" class so annotators do not force a guess. Pair each machine track with the interacting haul unit so you can derive cycle time, wait time and loads per hour.

Architecture choices affect what you buy. Two-pathway models such as SlowFast sample the same clip at different temporal rates [1], so frame rate and clip length in the delivered data must support both pathways; heavily compressed 5 fps archives will undercut them. If you need start and end times rather than clip labels, evaluate with temporal localization metrics (mAP at temporal IoU thresholds) as used in ActivityNet [2], and read our page on temporal action segmentation labels before writing the annotation spec.

Progress time-lapse and crew activity: weak labels from project records

Time-lapse progress data is most valuable when it is aligned to the project's own records. Daily logs, inspection reports, schedule activities and RFIs carry dates and locations that can act as weak labels for "formwork complete on level 3" or "curtain wall install started." Treat these as noisy: log entries can lag the work they describe, and schedule activity IDs drift between baseline and updates.

Crew activity video works best with a fixed task taxonomy agreed up front (for example, rebar tying, formwork, concrete placement, material handling, waiting). Construction tasks are procedural, so the approaches in procedural and instructional video datasets transfer well. Ask whether timestamps are camera-local or synced to a site clock, because multi-camera alignment fails quietly when clocks drift.

Who owns jobsite camera footage

Footage ownership on a construction project is usually set by contract, not by who mounted the camera. The candidates are the project owner, the general contractor, a specialty subcontractor (for equipment cameras) and the camera-service vendor that hosts the stream. Each may hold different rights, and the vendor's terms of service may limit what its customer can relicense.

Ask for the chain of documents: the owner-GC agreement clause on project records, the camera vendor contract, and any site-specific restrictions such as owner security policies on critical facilities. Footage often also contains signage, logos, adjacent buildings and radio audio, which are separate layers covered in rights layers in a video clip. Scraped jobsite video from public platforms does not come with any of these permissions.

Workers and the public in frame

Construction footage almost always captures subcontractor employees, inspectors, delivery drivers and passers-by, so privacy handling is part of the dataset spec, not an afterthought. Faces, name tags on hard hats, company decals on vests and license plates on haul trucks are the common identifiers.

Biometric statutes matter if anyone derives face geometry. As of October 2026, Illinois BIPA sets collection, retention and disclosure duties for biometric identifiers [3], and a 2024 amendment treats repeated collection from the same person as one violation [4]. Texas requires notice and consent before capturing a record of face geometry for a commercial purpose [5]. Large research corpora such as Ego4D collected video from consenting wearers and applied de-identification where needed [6], a useful benchmark for what to ask of wearable jobsite footage. For employee notice and audio rules, see recording workers on video for AI datasets and the broader de-identified data guide.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Construction video request checklist

Use this checklist to brief a supplier or broker; it forces the answers that usually stall a deal later.

Illustrative example: invented to show structure; it does not describe an available dataset.

An illustrative label record for an equipment-mounted clip might look like this:

{
  "clip_id": "site07_cam03_2026-04-14T07:00",
  "fps": 25,
  "equipment": [{"track_id": 12, "class": "excavator",
    "segments": [{"start_s": 0.0, "end_s": 6.4, "state": "dig"},
                 {"start_s": 6.4, "end_s": 9.1, "state": "swing_loaded"},
                 {"start_s": 9.1, "end_s": 11.0, "state": "dump"}],
    "paired_with": 31}],
  "persons_blurred": true,
  "weak_labels": {"daily_log_activity": "mass excavation, zone B"}
}

For container formats, manifests and transfer methods, see dataset delivery formats and transfer.

How SourceX approaches construction video requests

SourceX sources operational data from US companies on request; construction footage is not held in stock, and a request does not guarantee a match. Buyers describe the data they need rather than naming businesses, SourceX looks for US businesses that hold it, and every release is approved by the supplying company. SourceX does not source scraped web content or generic CCTV, so the useful requests are for jobsite video tied to an operation, labels and project records. Teams building construction-tech products can also review the construction industry buyer page and the construction management buyer page, or describe a video requirement to SourceX.

Each dataset is rights-reviewed for ownership and consents and delivered under a license defining records, uses, term and delivery. Personal details are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Delivery runs through private, access-controlled workflows only after an executed agreement and supplier approval.

Source construction site video for your model

If you need equipment cycles, crew activity or progress time-lapse footage with clear ownership, describe the capture type, labels and volume you need. SourceX will look for US businesses that hold matching footage and manage the process from assessment through licensing. Start a construction video request.