Vision Tracker
Track objects across video frames with Apple Vision, seeded by regions drawn in the viewport or fed from a detection node.
What it does
Vision Tracker follows objects across video frames from a seed region. Draw one or more seed regions over the image in the viewport, or feed in boxes from an upstream detection node, and each region is locked and followed frame to frame, reporting where it moved to, how confident the lock is, and a stable ID per tracked target.
Multiple targets are tracked at once, each with its own independent lock. This node runs on Apple's built-in vision tracking and is available on macOS only.
When to use it
A face, hand, product or other moving object followed through a video or camera feed without re-detecting it every frame
Position, scale or a trigger elsewhere in the graph driven by something moving on screen
Boxes from a detection node turned into a persistent, frame-to-frame track
| Parameter | Type | Default |
|---|---|---|
image | OneOf([FieldOf(Raster)]) | — |
seed_input | OneOf([Collection]) | — |
level | String | "fast" |
confidence_threshold | Number | 0.300 |
enabled | Boolean | true |
seed_regions | String | "" |
Gotchas
Seed regions take effect once. After a region is locked, that patch of image is followed on its own; moving or redrawing the seed region later does not nudge an existing track. Add a new seed region, or clear and restart, to retarget it.
Level trades speed for quality. Fast is meant for live, real-time feeds; Accurate spends more time per frame for a steadier lock, which suits pre-recorded footage better than a live camera.
A target that leaves frame, is occluded, or drops below Confidence Threshold has its track dropped rather than held onto indefinitely. Raise the threshold for stricter locks, or lower it to keep following through brief occlusions and motion blur at the risk of drifting onto the wrong thing.
Worked example
Wire a video or camera image into Vision Tracker's Image input.
In the viewport, draw a box around the object to follow. This becomes a seed region. Draw more boxes to track several objects at once.
Leave Level on Fast for a live feed, or switch to Accurate for recorded footage where the extra processing time is affordable.
Raise or lower Confidence Threshold until tracks survive the motion in your footage without drifting onto the wrong object.
Use the tracked output for position, box, and ID data downstream, or Positions for a simple list of centre points.