How to censor faces in dashcam footage for gdpr compliance
- Step 1Export the clip from the dashcam SD card — Copy the MP4/MOV/AVI off the card. Don't upload it to a cloud editor first — that's a transfer of personal data. Face pixelate runs locally and needs the Pro-Media tier (100 GB / 50 files); a typical 1–10 minute dashcam clip fits easily.
- Step 2Drop it onto the tool — Drag the file onto the drop zone. The detector warms up once (
Loading face detector). Detection then runs locally — no frames are transmitted. - Step 3Raise the sample rate for road speed — At driving speed, faces enter and leave frame fast. Raise
Sample rate (Hz)from the default 4 toward 8–12 so a pedestrian you pass is sampled enough times to form a track and get mosaiced. - Step 4Use a strong pixel size — For compliance, set
Pixel sizetoward 24–40 so detected faces are unrecognisable, not merely softened. The control clamps to 4–40. - Step 5Increase padding for partially-seen faces —
Padding (0–1), default 0.25 — raise to 0.35–0.4 so a face seen through a side window or at an angle is fully covered including the visible hairline and ear. - Step 6Run, then verify every identifiable person — Download the MP4 and scrub the whole clip. The short-range model misses distant/profile faces, so manually confirm each identifiable pedestrian/driver is covered; patch misses and license plates with the video redactor.
What this tool covers vs. what you must add for GDPR
Face pixelate handles detected faces only. A complete dashcam anonymisation usually needs more.
| Identifier | Covered here? | How to handle it |
|---|---|---|
| Near-camera pedestrian / driver face | Yes (auto-detected) | This tool, Pixel size 24–40 |
| Distant / profile / glare-obscured face | Often missed | Manual region in video redactor |
| License / number plate | No | video redactor — draw a region; see plate solution |
| Embedded GPS track in the file | No | metadata scrubber |
| Spoken identifying audio (names, addresses) | No (audio copied) | audio region muter |
Compliance-oriented settings
Recommended starting values for dashcam clips at road speed. Manual values — the UI has no presets.
| Scenario | Sample rate | Pixel size | Padding |
|---|---|---|---|
| Stopped at lights, close pedestrians | 4 | 30 | 0.35 |
| Driving past pedestrians | 10 | 28 | 0.4 |
| Oncoming drivers (front cam) | 12 | 28 | 0.4 |
| Car-park / low-speed manoeuvre | 6 | 30 | 0.35 |
Cookbook
Dashcam anonymisation scenarios. The tool is one part of the workflow — these notes show where it fits.
Junction incident clip for an insurer
You're submitting a clip to an insurer's portal. Pixelate the other driver and any pedestrians, but the audio (engine, your account) and the timestamp overlay should stay for the record.
Input: incident_front.mp4 Settings: Sample rate 10, Pixel size 28, Padding 0.4 Result: detected faces mosaiced, audio + timestamp overlay kept Then: redact the other car's plate in video-redactor
Clip headed for YouTube — full anonymisation pass
Publishing widely raises the bar. Do faces here, then plates and any missed distant faces in the redactor, then strip GPS metadata.
Step 1: face-pixelate (Sample rate 12, Pixel size 30, Padding 0.4) Step 2: /video-tools/video-redactor → plates + missed faces Step 3: /video-tools/metadata-scrubber → strip GPS track Result: publishable anonymised clip
Distant pedestrian not detected
A pedestrian 30 m down the road is too small for the short-range model and isn't mosaiced. There's no auto fix — cover them manually.
Symptom: small distant face still visible
Reason: short-range model targets near-camera faces
Fix: video-redactor → draw a region over the path they walk
(manual region, not auto-detected)Face through windscreen glare
An oncoming driver's face behind reflective glass may not detect. Raising the sample rate helps when they're briefly clear; otherwise redact manually.
Try first: Sample rate 12 (more chances to catch a clear frame) If still missed: video-redactor manual region over the windscreen
Plate is the real concern
If your DPA treats plates as personal data, faces alone aren't enough. Face pixelate does not detect plates — use the dedicated solution.
Face pixelate: faces only Plates: /video-tools/solutions/redact-license-plate-from-dashcam-video Do both for a clip with identifiable cars
Edge cases and what actually happens
Distant pedestrians not detected
May be missedThe MediaPipe short-range model is built for near-camera faces. A pedestrian far down the road, or a face that's only a few pixels tall, is frequently not detected and therefore not mosaiced. For GDPR you can't assume full coverage — review the whole clip and manually cover any identifiable distant person with the video redactor.
License plates are not touched
Not coveredThis tool detects and mosaics faces only. License plates — personal data under many DPAs when linkable to an individual — are completely ignored. Redact them with the video redactor; the plate solution walks it through.
Faces obscured by glare or angle missed
May be missedWindscreen reflections, harsh sun, and steep profiles defeat the detector. Raise Sample rate (Hz) to catch a clearer frame; if a face is never clear enough, it won't form a track. Cover it manually rather than assume it was anonymised.
No faces detected on a quiet road
Error: no facesIf the clip has no near-camera faces, you'll get No faces detected. That's expected for empty roads — but double-check it isn't because faces were too small to detect. If there are clearly people present, raise the sample rate or redact manually.
Embedded GPS track survives
Not removedMany dashcams write a GPS/speed track into the file. Pixelating faces doesn't touch it. Strip it with the metadata scrubber before publishing or sharing, otherwise the route/location is still in the file.
Spoken names / addresses in audio
Preserved (audio copied)Audio is stream-copied unchanged, so any spoken personal data (you reading a plate aloud, a name, an address) remains. Mute those segments with the audio region muter — see the interview-redaction solution.
Output re-encoded to MP4
By designThe result is always MP4 (libx264, CRF 20). The whole frame is re-encoded; at CRF 20 the un-redacted areas remain visually near-lossless for evidential clarity. The original container/codec is not preserved.
Tier gate
Tier requiredFace pixelate requires Pro-Media or higher. Free (1 GB / 1 file) and Pro (10 GB / 5 files) cannot run it. Pro-Media is 100 GB / 50 files; Developer 100 GB / unlimited. Limits are file-size/batch-count, so long dashcam clips are fine as long as the file fits.
Frequently asked questions
Is local pixelation enough to make dashcam footage GDPR-compliant?
Anonymising identifiable faces is a major part, but 'compliant' depends on your full processing. Plan for: distant faces the short-range model misses (cover manually), license plates (use the video redactor), the embedded GPS track (use the metadata scrubber), and spoken personal data in audio (use the audio region muter). This tool handles the detected-face layer. This is not legal advice — assess your own lawful basis.
Why is it important that nothing uploads?
The raw footage is itself personal data. Uploading it to a cloud anonymiser creates a transfer to a processor that needs its own legal basis and DPA. JAD runs detection and encoding entirely in your browser tab (TensorFlow.js + FFmpeg.wasm), so the personal-data clip never leaves your machine and you avoid that extra transfer.
Does it blur license plates too?
No. Face pixelate detects faces only — plates are ignored. Plates can be personal data when linkable to an individual, so redact them with the video redactor. The dashcam plate solution shows the workflow.
Why did it miss the pedestrian on the far side of the road?
The MediaPipe model used is the short-range variant, optimised for faces close to the camera. Small, distant faces (a handful of pixels tall) often aren't detected and so aren't mosaiced. Don't assume full coverage — scrub the clip and manually cover any identifiable distant person with the video redactor.
What sample rate should I use for footage shot while driving?
Raise it. At road speed faces pass quickly; the default 4 Hz may sample a passing pedestrian only once or twice. Use 10–12 Hz so they're caught enough times to form a track and get mosaiced. The trade-off is a longer detection stage.
How strong should the pixel size be for compliance?
For defensible anonymisation use Pixel size 24–40 so faces are reduced to large blocks with no recoverable structure. Lower values (4–12) leave a finer mosaic that may retain recognisable features on large close-up faces. The control is clamped to 4–40.
Does it keep the timestamp overlay and audio for my incident record?
Yes. The tool only edits pixels inside detected face boxes; a burned-in timestamp overlay elsewhere in the frame is untouched, and audio is stream-copied unchanged. So your evidential record (engine sounds, your spoken account, the on-screen time) is preserved — mute specific spoken personal data separately if needed.
What format do I get out, and is quality preserved?
An MP4 (H.264, yuv420p, CRF 20, +faststart). The whole frame is re-encoded but CRF 20 keeps non-redacted areas visually near-lossless, which matters for reading road details in an incident clip. The source container/codec is not retained.
Can I batch several clips from the same drive?
Pro-Media allows up to 50 files per batch and Developer is unlimited, both at 100 GB per file. So a folder of incident clips can be processed in one batch. Each file is detected and encoded independently with the same settings.
Does pixelating remove the dashcam's GPS data?
No. Pixelation only changes face pixels. Dashcams that embed a GPS/speed track leave it in the file. Run the metadata scrubber to strip GPS, device, and date metadata before you publish or share the clip.
What plan do I need?
Pro-Media or higher — face pixelate is gated on Free and Pro. Pro-Media: 100 GB / 50 files; Developer: 100 GB / unlimited. There's no minutes limit on video tools; the constraint is file size and batch count.
Is this legal advice?
No. This page explains what the tool does technically and where it fits in a privacy workflow. Whether a specific publication of dashcam footage is lawful depends on your jurisdiction, purpose, and lawful basis. Consult your DPO or a qualified adviser for compliance decisions.
Privacy first
Every JAD Video tool runs entirely in your browser via WebCodecs and FFmpeg (WebAssembly). Your video files never leave your device — verified by zero outbound network requests during processing.