Actions of bystanders in the capture range should not be treated as user input

What it is

Inside capture range there are not only operators but people passing, accompanying, cleaning, waving goodbye. Their motion is not user input. If the system takes a bystander’s wave as page or confirm, a non-participant has been written into the command channel. That is not “which of two people who both intend to operate should be heard”: the bystander never entered engagement and should not join the arbitration list. Rejecting bystander motion filters input out of traffic; it is not attributing among legitimate users.

Why it happens

The sensor has no “this is a passer-by” tag; bystander and operator are the same kind of pixel. A product with no engagement or identity gate treats any in-range motion that looks like the vocabulary as a candidate. A wave to family in the living room, a hand photographing an exhibit, a cleaner’s wipe, will hit templates. Even with a user already present, a larger, nearer bystander motion can steal the classifier’s attention. Filtering needs an engagement gate (only hands of someone who entered), a spatial gate (only bodies in the interaction volume facing the screen), or a temporal gate (a just-passing trajectory does not fire). With no gate the bystander is an anonymous second user who never appears in your test sample.

Studying it

While an operator finishes a task, have a confederate walk through, wave, point elsewhere. Record how often bystander motion becomes a command, whose head it is billed to, and whether the operator notices. Compare no gate, volume-only, engagement-only, and both. Negative samples should include social gestures aimed at the operator rather than the screen—that is the main shape of bystander motion. Do not substitute a second person instructed “please operate”; that measures multi-user arbitration.

Where it stops holding

On a stage or in a classroom the bystanders are an audience, and the product may want collective waves as input; that needs an explicit “crowd mode,” not the single-user filter. Infant or pet motion also hits templates; a filter that only assumes adult bodies will miss them. A truly private single-person room has low bystander risk and the gate can be looser. A headset almost only sees the wearer’s hands; the bystander problem moves to whether an outward-facing camera is still on.

Applying it

• By default accept only motion on a body that is engaged and facing the device. A passing trajectory must not issue a command even if the shape matches.
• Count “commands issued with no participant” as a bystander false-fire metric, separate from recognition accuracy.
• Accept in a real aisle or living-room cross-traffic: someone walks through and waves while an operation is in progress; the UI must not page. If it pages, a bystander entered the command channel.

Related

• Same group: C4.30.2 Bystander appearance and motion data also constitute privacy collection and need a handling policy · C4.30.3 The system needs to distinguish participants from third parties who are merely passing through · C4.30.4 Capture notices aimed at bystanders should be more conspicuous than those for users, because bystanders did not initiate the interaction
• Adjacent: C4.17 Input attribution and multi-user arbitration · C4.15 Conditions for entering engagement
• Search: bystander · false input · passerby