Skip to main content

Blur licence plates and other objects (track and censor)

ReelBolt can find an object in your footage, follow it as it moves, and obscure it — the typical use is blurring licence plates, but it works for any flat thing you can name: a shop sign, a name badge, a screen showing something private. No green screen is needed.

It takes two steps:

  1. VideoAnalyze finds and follows the objects. You type what to look for ("licence plate"); a vision AI finds every instance on frames spread through the clip, and a tracker then follows each one frame by frame — through camera moves, zooms and tilts.
  2. VideoCompile obscures every tracked object whose name you list — blurred, pixelated or solid black — and keeps following it in the finished video.

What you need to censor objects​

Before you start:

  • A vision AI provider set up by your administrator (any provider with the Vision capability). It is only used to find the objects; following them uses no AI.
  • Your footage, uploaded to the project.

Cost: each "look" at a frame is one or two calls to the vision provider. By default ReelBolt looks once a second, at most 12 times per clip.

Build the track-and-censor workflow​

Step 1 — VideoAnalyze. Pick your video as the source. Under Track objects, type what to look for, for example licence plate, and press Enter. In JSON:

{
  "version": 1,
  "source": {"kind": "ProjectFile", "projectFileId": "PUT-THE-FILE-ID-HERE"},
  "transcription": "Off",
  "trackObjects": [{"label": "licence plate"}],
  "objectDetectEverySec": 1.0,
  "maxObjectDetectFrames": 12
}
  • trackObjects — what to find. Use a short, plain description; you can list several.
  • objectDetectEverySec — how often ReelBolt looks again, so cars that drive in later are caught too. Lower is more thorough and costs more lookups.
  • maxObjectDetectFrames — the most lookups per clip. On a long clip they are spread evenly over the whole clip.
  • objectTrackMaxSide (default 960) — the resolution the tracker works at. Raise it (up to 1920) for small, distant plates.

Step 2 — VideoCompile. Switch on Keep every source whole (no edit) if you don't want the video cut, and under Censor tracked objects type the same label. In JSON:

{
  "version": 1,
  "decision": {"from": "Previous"},
  "analysisStepOrder": 1,
  "keepWholeSources": true,
  "censorLabels": ["licence plate"],
  "censorStyle": "Blur",
  "censorPadding": 0.25,
  "censorHoldSec": 0.4
}
  • censorLabels — which tracked objects to obscure, by the name you gave them in step 1. "*" obscures everything that was tracked.
  • censorStyle — "Blur", "Pixelate", or "Fill" (solid black; the only style that leaves nothing of the original behind).
  • censorPadding — extra cover on every side, as a fraction of the object's longer side (0.25 = a quarter of a plate's width above, below and either side), to cover small detection and tracking errors.
  • censorHoldSec — extra seconds of cover before an object is first tracked and after it is last tracked.

You can also put an edit in between: keep a VideoStoryEditor step as step 2 and point the compile step's decision at it. Censoring follows the object into every part of the finished video where it appears, however the edit cut it up.

Check the censored result​

  • Step 1's result shows meta.objectTracking: how many objects were tracked (tracks), how many vision lookups were made, and — if something went wrong — degraded: true with a reason (for example "no Vision provider is configured"). Progress messages name each object as it is tracked, the settings the tracker chose for it, and its confidence.
  • Step 2's result has a censor section listing every obscured stretch, and an unmatchedLabels list. Anything in unmatchedLabels was not obscured anywhere — no object with that name was tracked. Check this list every time.
  • If finding objects went wrong (no vision provider, a failed lookup, too many objects for maxObjectTracks), step 2 stops with an error instead of making a video, because the result might show what you wanted hidden. Fix the cause and run again.
  • The Editor shows each censored object as a red dashed outline over the video, with its name and confidence, on its own Censored track in the timeline.

Limits of automatic censoring​

Read this before publishing anything where privacy matters:

  • The vision AI can miss objects, especially small, blurry or partly hidden ones. Always watch the result.
  • Between lookups, a newly appearing object is only found at the next lookup. With the default of one look per second, a plate that drives in can be visible for up to a second before it is obscured. Lower objectDetectEverySec for busy scenes.
  • A tracked object that gets fully covered (a person walks in front) ends its track; the next lookup starts a new one. censorHoldSec covers short gaps.
  • Blur and pixelation are not irreversible. For anything sensitive, use "censorStyle": "Fill".
  • The tracker follows flat things best. A face turning away, or an object that bends, is followed less precisely; raise censorPadding to compensate.

Placing a region by hand instead of searching​

When you know exactly where the object is at one moment, you can skip the vision AI: give the target a region (its four corners as fractions of the frame, top-left, top-right, bottom-left, bottom-right) and the time it is there. The tracker follows it from that moment, forward and backward, with no AI call:

"trackObjects": [
  {
    "label": "shop sign",
    "atSec": 2.0,
    "region": {"x0": 0.62, "y0": 0.18, "x1": 0.80, "y1": 0.18,
               "x2": 0.62, "y2": 0.27, "x3": 0.80, "y3": 0.27}
  }
]

The assistant can fill this in for you: tell it what to obscure and roughly where.