Short answer
PPE recognition data means images of hard hats, vests, gloves, eyewear, hearing protection and respirators, both on people and off them, across sites and lighting. Operating businesses may hold such images, but they show workers, which brings consent, privacy and misuse questions. A PPE detector can inform a robot's behavior around people; it does not on its own make a robot safe. This is not legal advice.
The buyer's problem
A humanoid working in a shared space has to behave differently around people in different roles. One reason to recognize PPE is as a cue about roles or zones, treated as a hint for the robot's own behavior and never as proof that someone is trained or authorized. Another is to check that its own tasks, such as handing over tools or restocking a protective-gear station, use the right items. A third is evaluation: testing whether the robot's perception correctly recognizes gear in the scenes it will work in. That is different from monitoring whether individual workers comply with site rules, which is a separate use with its own consent and labor questions.
Which of these is your purpose changes what data you need. Detecting a hard hat on a head from forty feet is a different problem from classifying glove type in a close-up. This page covers sourcing the images. For safety signage and labels, see workplace signage and safety labels, and for the overall cluster, the product and object data hub.
What the data should vary
Ask for variation along the dimensions where PPE detectors typically break. Whether a given supplier covers them is a diligence question.
- Worn versus not worn. A hat on a head, a hat on a bench and a hat hung on a hook are three different states. Labeling "present" without "worn" produces false reassurance. See object state recognition.
- Worn correctly. Chin straps, vest closure, goggles over the eyes versus on the forehead, respirator seal.
- Item variety. Colors by role, brands, reflective patterns, regional norms, disposable versus reusable.
- Viewpoint and distance. Overhead cameras, eye-level cameras and robot head height give very different pixels. See small object resolution requirements.
- Occlusion and crowding. Workers behind equipment, bent over, in groups. See cluttered scenes for recognition.
- Lighting and weather. Backlit docks, night shifts, welding glare, dust, rain.
- Non-compliant and ambiguous cases. Caps resembling hard hats, hi-vis clothing that is not a vest, partial gear.
A note on sources
OSHA's ergonomics overview describes musculoskeletal disorders and frames an ergonomic process as an ongoing function within a workplace safety program (OSHA). It does not address PPE recognition or computer vision. It is cited here only for the framing that safety is ongoing and managed, which is a reason to treat a PPE dataset as one input to a broader safety case, not a result in itself. This page has no task-specific published source for PPE detection and cites no PPE regulation or standard; claims about detector behavior are stated as reasoning and questions, and which PPE a site requires is for the site's own safety program and advisers to define.
Illustrative annotation schema
Illustrative only; synthetic values and arbitrary choices that a buyer would replace.
| Field | Illustrative value | Note |
|---|---|---|
item_type | hard_hat | Class set decided by the buyer |
wear_state | worn_correct / worn_incorrect / carried / stowed | Presence alone is not enough |
person_visible | partial | Identifiable faces drive consent questions |
distance_band | mid | Defined by apparent pixel height of the item |
occlusion | moderate | Labeler-assigned, with guidelines |
lighting | backlit | Condition tag |
site_id | site_031 | For held-out-by-site evaluation |
Define bands such as mid in pixels, not words, in your own spec. The values above carry no recommended thresholds.
Where the images come from
Operating businesses with safety programs hold site photos, audit images, training material and incident documentation. Construction, manufacturing, warehousing, food processing and utilities can produce such imagery; whether a given business retained it, and may license it, is a diligence question. The material is usually realistic and diverse. It is also full of identifiable people.
Safety training providers and PPE manufacturers hold staged images and catalog photos. These are clean and rights-clearer, but depict models in ideal poses.
Fixed-camera archives from site monitoring capture many frames cheaply. They also capture everyone who walked past, and a camera installed for security or safety may have been set up under notices that did not mention training.
A purpose-built capture, with consenting participants wearing and misusing a defined PPE set, offers control and cleaner consent, at the cost of realism and effort. Compare the routes in existing assets, new collection or specialist partners.
People in frame: permissions and provenance
PPE images usually show people, so permission is typically the main gating item. Ask what notice workers received, whether employment agreements or union terms address recordings, and who inside the business can authorize a license. Ask whether any state-law rules on recording or biometric identifiers apply to the sites and the way images were produced. Blurring faces may reduce exposure but does not settle whether a business may license the images. For the worker-side questions see worker consent and identifiable people. This is not legal advice.
Redaction also affects training value. Blurred heads can remove the very region where a hard hat or goggles sit. If the supplier plans to redact, ask which regions and whether PPE remains visible.
What PPE recognition data cannot support
A PPE detector reports what a model believes it sees. It does not establish that a person is trained, authorized, in a safe position or aware of the robot. Recognition does not give a humanoid safe motion planning around people, which needs its own controls, validation and site risk assessment. It also gives no data about how to hand an object to a gloved hand or work next to someone in a respirator; those behaviors need robot or teleoperation data and human-robot interaction testing.
Further limits:
- Distribution shift. A detector tuned on one industry's gear may fail on another's.
- Label ambiguity. Whether a vest "counts" depends on a site's rule, not a pixel property.
- Evaluation integrity. Hold out sites and, where appropriate, individuals, so the model is not tested on people it was trained on.
- Misuse risk. A dataset built for robot perception can be repurposed for surveillance or compliance monitoring of individual workers. Ask the supplier whether workers were told about such uses, consider whether the license should exclude identifying or tracking individuals, and review this with counsel. This is not legal advice.
For how workplace incidents are documented, which sometimes yields PPE-related imagery under tight controls, see near-miss and incident reports.
Questions to ask suppliers
- Which industries, sites and PPE types are represented, and how many distinct workers?
- Were images taken for safety audit, training, security or marketing, and under what notices?
- Are wear-state and correct-use labels present, and who assigned them?
- How are faces and identifiers handled, and what does redaction do to the PPE region?
- What camera heights, distances and lighting conditions appear?
- Are non-compliant and ambiguous cases present or filtered out?
- Who can authorize licensing: the business, the employees, a contractor or a union?
- Do workers or contracts restrict reuse for model training?
- Can a whole site be held out for evaluation?
Discuss your requirements
If your perception work needs PPE imagery, describe the item types, viewpoints and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match; SourceX follows up about potential sources. Start at the humanoid training data guide for the broader routes, and see the SourceX trust page for how data handling is described.
Discuss your humanoid training-data requirements with SourceX
Sources
- Ergonomics - Overview — OSHA's framing of safety as an ongoing workplace process; not a source on PPE recognition.