Privacy and preparation
How to preserve motion while reducing worker identifiability in video
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Preserve motion while reducing worker identifiability by naming the motion signals a robotics buyer needs, applying the lightest transformation that removes other identity cues, and testing each candidate on the same clips for both motion utility and re-identification. Face blurring alone rarely settles it, because gait and movement style can still identify a person, so keypoints, crops and generative replacement need their own recognition tests.
Video and sensors
Preserve motion
Key takeaways
- Write down the motion signals the buyer will train on before choosing any transformation.
- Face blurring is a first step; gait and movement style can still identify a worker.
- Measure motion utility and identifiability on the same clips, against untouched originals, with a decision rule set in advance.
- Check that fumbles, regrasps and handoffs survive the transformation, not just average pose scores.
- Small stations, visible hand marks and requests for raw files are reasons to narrow scope or bring in a privacy reviewer.
How do you keep motion data while making workers harder to identify?
Start from the motion the robotics team will actually use, then apply the lightest transformation that removes identity cues outside that signal, and test both sides on the same clips. One test checks that hand and body pose, contact and timing survive the change; the other checks whether a person can still be recognized from what is left.
Face blurring is usually the first step, not the last one. Research on virtual reality users and on gait shows that the way a person moves can single them out, so a site that keeps full-body motion has to treat motion itself as a possible identifier. The plan below gives a business a repeatable way to choose a transformation, record why it was chosen and show a buyer the evidence.
This page is about the transformation decision. Notice, consent and opt-out mechanics come before any of it and are covered in the related guides linked at the end.
Which parts of workplace footage carry identity, and which carry the task?
Identity and task signal overlap less than people assume, and the overlap is where the hard choices sit. A robot learning team usually needs hands, tools, objects, contact events and the timing of each step; it rarely needs a face, a voice or a name badge.
The overlap is motion. Nair and colleagues report that more than 50,000 virtual reality users could be identified across sessions from head and hand movement alone, and gait recognition research such as GaitSet identifies people from walking silhouettes. Ego4D, a large egocentric video dataset, combined automated face and plate blurring with manual review and blurred other personal details such as written names by hand.
| Cue in the footage | Needed for robot learning? | Typical handling |
|---|---|---|
| Face and head | Rarely; sometimes gaze direction | Blur or replace; keep a gaze label if the buyer needs it |
| Voice and speech | Rarely; sometimes spoken step names | Strip audio or keep a reviewed transcript without names |
| Badges, tattoos, printed names, screens | No | Mask by hand-checked detection or exclude the frames |
| Body shape and clothing | Partly, for reach and posture | Replace the body, standardize clothing or keep keypoints only |
| Gait and movement style | Often, it is the task | Cannot be removed without changing the motion; limit and test |
| Hands, skin marks, rings | Yes, for grasp and contact | Gloves on capture day, or targeted masking with a utility retest |
| File metadata, timestamps, camera IDs | Relative timing only | Strip absolute dates and device serials; keep frame times |
What each transformation keeps and what it gives away
Every transformation trades some motion fidelity for some identity reduction, and the trade depends on camera angle and task. A study of industrial image and video anonymization compared pixelation, blurring and DeepPrivacy2, a generative full-body replacement method: pixelation made body detectors fail, blurring reduced pose and action recognition quality but stayed usable, and DeepPrivacy2 gave better body and face pose than blurring but was weaker for precise hand pose and was temporally inconsistent.
Detector-based blurring also misses things. The EgoBlur project, which blurs faces and license plates in Project Aria recordings, notes that some plates in output video may remain unblurred because of low detection confidence and that its score threshold can be lowered to add low-confidence detections. Any automated pass therefore needs a sampled human audit.
| Transformation | Motion it keeps | Identity it can leave | Main failure to test |
|---|---|---|---|
| Face blur (detector plus blur) | Full body, hands, timing | Gait, body shape, clothing, missed faces | Missed detections in profile, occlusion, low light |
| Pixelation of people | Coarse location only | Low pixelation can still reveal people | Pose detectors stop finding bodies |
| Generative face or body replacement | Body and face pose, mostly | Gait and movement style | Hand pose drift and frame-to-frame flicker |
| Keypoints or skeleton only | Joint trajectories and timing | Gait and motion signature | Object state and contact lost with the pixels |
| Crop to hands and work surface | Grasp, tool use, contact | Skin marks, rings, sleeves | Reach and posture context removed |
| Audio strip or reviewed transcript | All visual motion | Names said on screen text | Spoken step cues lost |
| Metadata strip and relative time | Frame timing | Linkage through other records | Sync between cameras broken |
Can capture-day choices reduce how much transformation is needed?
Yes, and they are usually cheaper than heavier post-processing. Every identity cue that never reaches the sensor is one the transformation does not have to remove, and one the evaluation does not have to test.
At a glance
Can capture-day choices reduce how much transformation is needed
The tradeoff is that capture choices also shape the motion. A camera placed to avoid faces may also lose the shoulder angle the buyer needs, and uniform gloves change grip. Agree the capture plan against the motion contract before the first session, then treat it as part of the method the buyer receives.
- Camera angle: mount fixed cameras above or behind the work surface so faces are rarely in frame, if the buyer does not need full posture.
- Head-mounted cameras: egocentric views show the wearer's hands but rarely the wearer's face, while coworkers in view still need handling.
- Clothing and gloves: offer plain sleeves and standard gloves, and record which sessions used them so grip analysis can account for it.
- Badges and screens: turn badges around and keep screens with customer data out of the capture zone.
- Audio: disable microphones unless spoken step names are part of the contract.
- Device settings: switch off location tagging and record only relative timestamps where the hardware allows it.
The transformation tradeoff evaluation plan
The plan runs each candidate transformation over the same stratified test clips, measures motion utility against the untouched originals and measures identifiability with fixed attacks, then applies a decision rule written before anyone sees the scores. Keep the originals under restricted access while the test runs, and decide in advance whether they are deleted afterward.
Use it as is or adapt the metrics to the buyer's task. The plan reports what was measured; it does not promise that a transformation makes footage anonymous.
| Step | What to measure | How | Decision it feeds |
|---|---|---|---|
| 1. Motion contract | Signals the buyer will train on | Written list: joints, contact events, objects, frame rate | Which cues must survive |
| 2. Test clip set | Coverage of angles, light, PPE, crowding | Stratified sample of clips from every camera and shift | Whether results generalize across the site |
| 3. Baseline extraction | Pose, contact and action labels on originals | Same models and labelers used later | Reference for every utility score |
| 4. Utility after transformation | Keypoint shift, detection drop, label agreement, flicker | Compare outputs on transformed versus original clips | Whether the motion contract still holds |
| 5. Detection audit | Faces, badges and screens left visible | Human review of a fixed share of frames per clip | Whether the automated pass needs rework |
| 6. Recognition attacks | Face match, gait match, coworker recognition | Enrolled gallery of consenting staff; reviewers who know the team | Residual identifiability by cue |
| 7. Linkage test | Matching transformed clips to rosters or originals | Try to join on time, station and camera | Which metadata must be cut |
| 8. Pre-set decision rule | Pass, revise or exclude per clip group | Thresholds written before step 4 | Final transformation choice |
| 9. Record for the buyer | Method, versions, limits, failures | Short method note shipped with the data | What the buyer is told about residual risk |
Why a blurred face does not settle whether the worker is identifiable
Removing appearance does not remove motion identity. A study in Sensors, published online in December 2025, applied DeepPrivacy2 to industrial CCTV-style recordings and found that full-body anonymization removed visual appearance while gait identity stayed largely intact; the authors conclude that pose-driven anonymization alone is not enough.
Some laws reach movement and geometry directly. The GDPR defines biometric data as personal data resulting from specific technical processing of physical, physiological or behavioral characteristics that allows or confirms unique identification, and the Illinois Biometric Information Privacy Act lists a scan of hand or face geometry as a biometric identifier. Whether a particular keypoint export falls under either is a question for counsel on the specific facts.
For the evaluation plan this means step 6 cannot stop at a face recognizer. A small team at one site is the hardest case, because coworkers can often recognize a colleague from posture and pace alone.
Worked example: an illustrative packing-station capture
Illustrative only. A distributor films a packing station with two fixed cameras and one head-mounted camera for a buyer that wants hand-object contact and upper-body pose. The motion contract lists both hands, wrists, elbows and shoulders, the moment of each grasp and release, and the box and item positions.
The team draws a test set of 120 clips across both shifts and all three cameras and compares three candidates: face blur only, face blur plus generative body replacement, and keypoints plus a hand crop. In this example, face blur keeps the motion contract but fails the coworker recognition test for most clips; body replacement passes recognition for clothing and build but loses wrist precision during fast regrasps; keypoints plus the hand crop keeps the contract and passes recognition except on one station where only two people work.
The exception: one worker has a distinctive forearm tattoo that stays visible in the hand crop. Masking it moves the wrist keypoints, so the team chooses long sleeves on capture days for anyone who wants them and drops the earlier clips of that worker rather than ship a crop that names them. The two-person station is excluded from delivery. The decision and the failed candidates go into the method note.
Which variation and recovery moments disappear after the transformation?
The transformation can quietly remove the behavior a robotics team values most. Fumbles, regrasps, small corrections and hesitations are short and low-contrast, and they are exactly where blurring, replacement flicker or a tight crop do the most damage.
Label these moments on the originals before transforming, then check that each one is still visible and correctly timed afterward. If recovery events drop out at a higher rate than routine steps, the transformation is biasing the dataset toward clean runs even when average pose scores look fine.
- Regrasp and drop events: still visible and labeled at the same frame range.
- Gaze or head turn before a correction: kept as a label if the face is removed.
- Two-person handoffs: both hands tracked without identity swaps between workers.
- Tool changes and PPE adjustments: not masked as if they were badges.
- Low-light and occluded segments: not dropped silently by the detector.
Signals that the footage should be narrowed or held for privacy review
Some results mean the scope should shrink rather than the transformation get heavier. Hold the clip group, and bring in a workplace privacy reviewer and counsel, when any of these appear.
- A station has so few workers that motion alone points to one person.
- Hands carry marks that cannot be covered or masked without breaking the motion contract.
- Audio or screens contain names, customer details or other third-party data.
- Customers, visitors or contractors appear and were not part of the consent process.
- The buyer asks for unblurred originals or raw device files with serial numbers.
- Keypoint exports could be treated as biometric data in a jurisdiction where the site or workers are located.
- The recognition attack in step 6 cannot be run because nobody consented to an enrolled gallery.
Where does transformed footage fit in a licensing review?
A transformation result answers how identifiable the footage is; it does not answer whether the business may license it. Ownership of the footage, worker permissions and the contract terms are separate questions, and a de-identified clip still needs all three.
Rights review, contracting, preparation and delivery all run through SourceX when a company licenses selected footage, and an upfront assessment, without any data upload, settles whether the company qualifies. The business signs off on scope and terms before any footage moves, and because SourceX performs the de-identification through approved processing routes, a method note like the one above becomes part of that review.
Frequently asked questions
Is face blurring enough to anonymize workplace video?
Usually not on its own. Gait and movement style can identify people, and detector-based blurring misses some faces in profile, occlusion or low light. Treat face blurring as one step and test what remains.
Can skeleton or keypoint data identify a worker?
It can. Research on motion data shows people can be singled out from head and hand movement or from gait, so keypoint exports need their own recognition test. Whether such data counts as biometric under a given law is a question for counsel.
Does blurring hurt robot learning data?
It can reduce pose and action recognition quality, and published industrial tests found pixelation made body detectors fail. The effect depends on camera angle and the task, which is why the plan measures utility on the buyer's own signals.
Should the original unblurred footage be kept after transformation?
Keep it only as long as the evaluation and any agreed audit need it, under restricted access, and record the deletion date in advance. Keeping originals indefinitely keeps the identification risk the transformation was meant to reduce.
Who should run the re-identification test?
Someone independent of the capture team, using a gallery of staff who agreed to take part. Include reviewers who know the workers, because coworker recognition is often the strongest attack at a single site.
Sources
- Nair et al. report that more than 50,000 virtual reality users could be uniquely identified across sessions from head and hand motion data. Source
- GaitSet treats gait as a set of silhouette frames and learns identity information from it for cross-view gait recognition. Source
- Ego4D de-identified non-consented footage with automated face and plate detection plus manual review, and blurred other personal details such as written names by hand. Source
- A study of industrial video anonymization found pixelation made body detectors fail, blurring reduced pose and action quality, and DeepPrivacy2 preserved body and face pose better but was weaker for hand pose and temporally inconsistent. Source
- EgoBlur detects and blurs faces and license plates in Aria images and video, and its documentation notes that low-confidence detections may remain unblurred unless the threshold is lowered. Source
- A Sensors study published online in December 2025 (vol. 26, no. 1) found full-body anonymization with DeepPrivacy2 removed appearance but left gait identity largely intact in industrial monitoring footage. Source
- GDPR Article 4(14) defines biometric data as personal data from specific technical processing of physical, physiological or behavioral characteristics that allows or confirms unique identification. Source
- The Illinois Biometric Information Privacy Act defines biometric identifier to include a scan of hand or face geometry and excludes photographs. Source
See if your company qualifies
A short company assessment. No data uploads are needed.