AI inference (camera analytics)

With a camera attached, a screen can run AI inference to measure who is watching — audience attention and demographics — and feed that back as insights. It's how impressBox powers attention/insight analytics and audience-aware behaviour.

Modes

• Image mode — the device periodically captures and analyses still frames.

• Stream mode — the device streams to a real-time analysis backend (WebRTC).

What you configure

• Enable inference on the device and pick the mode.

• Processing interval — how often frames are analysed (or as fast as possible).

• Image size for processing.

• Camera geometry — field of view, camera tilt/position, and the screen's width/height, so detections align with the display.

• Black-stream watchdog — restart if the camera stream goes black.

• (Stream mode) Chunk length — how results are grouped over time.

Typical uses

• Measure impressions, dwell time, and attention for proof-of-attention reporting.

• Audience-aware content (adapt what plays to who's in front of the screen).

• Combine with content evaluation against a device for placement decisions (see Previewing & testing on a real device).

Privacy & setup notes

Camera analytics must respect local privacy rules and signage/consent requirements. Quotas for inference and attention-insight screenshots are set in account settings (see User settings). Set this up with impressBox so the camera, backend, and geometry are configured correctly.