When several people share the sensing volume, the system must decide who owns the current action

What it is

A mid-air sensor observes motion in a volume; it does not receive a pre-labeled “command from person X.” When two or more people occupy the sensing volume, the system has to perform input attribution: decide which person produced this wave, pinch, or point. Without that step, even a perfect recognizer is only classifying anonymous trajectories. This is not the bystander problem of “should a passer-by count as a user at all.” It is the problem of binding one action to one person among co-present people who could each be operators.

Why it happens

Depth cameras, stereo vision, and millimeter-wave radar emit joints or motion blobs per frame. With several bodies present, the tracker keeps multiple trajectories, but IDs swap or break under occlusion, crossing paths, and people entering or leaving the frame. If the product maps “a valid gesture was detected” straight onto a command, it is implicitly assuming a single legitimate sender. In a living-room TV, a museum kiosk, or a sterile imaging console, a second person’s conversational gesture, hair adjustment, or pointing-to-a-companion can overlap the first person’s action in time. Attribution turns “who is visible” into “who is allowed to issue this frame’s command”; it is not an optional extra on top of recognition.

Studying it

Two (or three) people stand in front of the same depth camera. One performs the target gesture; the other talks, points at an exhibit, or walks. Independent variables include stance distance, body overlap, and whether crossing through is allowed; dependent measures include correct-attribution rate, identity-swap count, and commands executed for the wrong person. Wearable markers or hand-labeled skeleton identity serve as ground truth, because consumer skeleton IDs are themselves unstable. Measuring only single-user recognition folds attribution failures into “the gesture was not recognized” and hides the actual error class.

Where it stops holding

A head-worn display or a wearable that only tracks the wearer’s hands never feeds a bystander’s hand into the same joint set, so the hardware envelope removes the attribution problem; multi-user arbitration is unnecessary. Conversely, a public display that “serves one person at a time” still needs explicit attribution whenever two bodies can appear in the sensor’s field of view. Child height, wheelchair use, and a person carrying an infant change skeleton count and occlusion patterns; attribution rates measured on standing adults do not transfer unchanged.

Applying it

• Write a hard limit on how many people may be simultaneously visible, and specify whether overflow rejects new input, freezes the current session, or enters an explicit multi-user mode.
• Show attribution to the operator: the person currently treated as sender needs a stable highlight, index, or audio cue, rather than learning who was heard only after a command fires.
• Record two-person tasks in the actual venue. Count cases where the gesture class was correct but billed to the wrong person, separately from recognition errors.

Related

• Same group: C4.17.2 Attribution depends on identity features, not on the motion itself · C4.17.3 Misattribution treats one person's motion as another person's command · C4.17.4 Arbitration rules must be predictable to users, such as first-come or nearest-person
• Adjacent: C4.30 Bystanders and third parties · C4.15 Conditions for entering engagement
• Search: input attribution · multi-user sensing · skeleton identity