Workplace Safety Video Data: Near-Misses, Unsafe Acts and Incident Footage

Why near-miss footage is the realistic positive class

Near-misses, not injuries, are where most usable positives come from, because recordable incidents are rare at any single site. In OSHA usage, a near miss is an event with no injury or damage where a slight shift in time or position could have caused harm, and most safety programs encourage reporting them even where no rule requires it. Sites that run a near-miss reporting program therefore hold far more event records than injury records, and each report carries a timestamp, location and narrative you can join to camera footage.

Unsafe acts sit one level below near-misses: a worker stepping over a conveyor guard, riding a forklift fork, or standing under a suspended load without an event occurring. These come from behavior-based safety observations and supervisor walk-downs rather than incident systems. A training set that mixes all three tiers needs an explicit severity field, or your detector will learn to treat a missing hard hat and a struck-by event as the same class.

Plan for imbalance from the start. Ask suppliers how many camera-hours sit behind each labeled event, because the negative footage between events is what calibrates false-alarm rates on a live site.

Labels come from safety records, not from the video alone

The most defensible labels come from the operator's investigation trail, joined to video by time and camera. Under OSHA's Part 1904 recordkeeping rule (29 CFR 1904.29 covers the forms), employers record each recordable case on the OSHA 300 Log and complete a 301 Incident Report or equivalent within seven calendar days [1]. A 301 narrative (what the employee was doing, what happened, the object or substance involved) gives you a ground-truth event description that an annotator then localizes in time.

Label definitions must follow the site's own rules, not a generic ontology. "PPE violation" means nothing until you know whether the zone required cut-resistant gloves, a face shield or high-visibility vests, and "exclusion zone breach" depends on painted floor markings, light curtains or the site's traffic plan. For forklift work, many operators map rules to 29 CFR 1910.178, the general-industry standard for powered industrial trucks, and OSHA's interpretation letters show how those provisions apply in practice [2]. Ask for the site rulebook alongside the footage, and record which version applied at capture time.

For temporal labels, specify onset, peak and resolution timestamps rather than a single clip label; the guide to temporal action segmentation labels covers the boundary conventions. The written records themselves are a separate purchase, covered in HSE incident and near-miss reports for AI.

Illustrative event record for a near-miss clip

A per-event record should link the clip, the source record, the governing site rule and the redaction status, so reviewers can audit each label.

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

Litigation holds and insurer claims can lock incident footage

Footage of an actual injury is often the hardest category to license, because it may be evidence. Once a claim, workers' compensation dispute or lawsuit is reasonably anticipated, the employer's counsel typically places the clip under a litigation hold, and the insurer may have its own copy and claim file. Releasing or editing held footage can create spoliation risk for the supplier, so expect a legal-hold check per clip and exclusions for open matters.

Privacy rules travel with the records too. OSHA lets employers log certain sensitive cases as "privacy case" with a separate confidential list of names [1]; a video clip of that same event can defeat that protection if it is not redacted. Treat any join between footage and 300 or 301 data as health-adjacent information and limit fields to what the label needs.

Staged reenactments versus real events

Staged reenactments buy rights clarity at the cost of realism. Actors who sign releases, on a closed site with planned camera angles, give you clean consent and no legal holds, but their motion, hesitation and occlusion patterns differ from a tired worker at the end of a shift. Real events carry the opposite trade: authentic dynamics, but the people recorded never agreed to an AI use.

A practical split is to train on real near-misses and unsafe acts, use reenactments to fill rare classes such as struck-by or caught-between, and keep a held-out evaluation set of real events only. Flag every staged clip in the metadata so evaluation never mixes the two. For consent mechanics when filming employees, see recording workers on video for AI datasets.

Faces, biometrics and the rights in each clip

Workplace safety footage nearly always shows identifiable workers, so treat biometric statutes as part of the purchase review. Illinois BIPA section 15 sets written-release, retention and disclosure rules for biometric identifiers held by private entities [3], and the 2024 amendment treats repeated collection from the same person by the same method as one violation [4]. Washington's biometric statute excludes photographs, video recordings and data generated from them from its definition of a biometric identifier [5]. Raw footage alone may fall outside these definitions, so the answer depends on what you extract, such as face geometry, embeddings or gait templates.

Most unsafe-act and PPE models do not need identity, so ask for face blurring before delivery and confirm that no identity features were derived. Badges, helmet names and vehicle numbers also identify people. The broader rights picture is covered in rights layers in a video clip and biometric data in AI training.

How this differs from PPE image sets and routine task video

Safety video is a different purchase from still PPE images and from routine operations footage. Public construction-safety sets are mostly still frames: one example offers 1,214 manually labeled images from four static site cameras with two posture classes [6]. Stills train helmet and vest detectors but cannot teach approach speed, line-of-fire or the seconds before a near-miss.

Routine task video, such as warehouse operations footage or manufacturing assembly video, is captured for productivity and process steps and rarely carries safety-event labels. Fleet and driving safety have their own page on dashcam video datasets. If you are scoping by sector, the construction buyers and manufacturing buyers pages describe related requests, and inspection reports can add hazard labels to fixed-site footage. Start from the video data hub for the full cluster.

Questions to put to a supplier of safety footage

Your diligence pack should document source systems, label provenance and release decisions per clip, in the spirit of structured dataset documentation such as Data Cards [7]. Ask for:

• The safety record systems used for labels (near-miss reports, observation cards, OSHA 300 and 301 or equivalents) and how clips were joined to them.
• The site rulebook and its version history for each labeled zone or PPE requirement.
• Camera-hours per labeled event, plus negative footage sampled from the same cameras.
• A per-clip legal-hold and claims check, with open matters excluded.
• Face and identifier redaction method, and confirmation that no biometric templates were created.
• Whether any clips are staged, with a flag in metadata.
• Allowed uses in the license: training, evaluation, or both.

On the SourceX side, every dataset is rights-reviewed for ownership and consents, personal details are removed or replaced before delivery with the method recorded and a sample checked, and no method is perfect. SourceX does not source generic CCTV, and diligence materials covering source, rights, preparation and allowed use are prepared per dataset. You can describe the footage and records you need without naming suppliers.

Sourcing workplace safety video datasets with SourceX

SourceX looks for US businesses that hold the operational data you describe, including recordings of hands-on work, and every release is approved by the supplying company. Datasets are sourced on request rather than held in stock, so a request does not guarantee a match, and nothing is contracted until a supplier agrees. Describe the safety footage you need.

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