HeadSense: Head Orientation Tracking for AI Wearables using Ultra-Wideband Ranging
AI wearables such as smart glasses, pins, pendants, and earbuds rely on head gestures for interaction, yet they lack a reliable, continuous measure of where the head points relative to the torso. That signal gives AI assistants temporal attention context: it lets them infer sustained visual attention, resolve spatial commands, and filter out incidental head movements. IMU-based tracking drifts over time, and magnetometer correction fails indoors near ferrous materials and the speaker magnets common in wearables.
HeadSense is a necklace-like wearable that uses ultra-wideband (UWB) radios to track head orientation relative to the torso with bounded, non-accumulating error. A wired reference path between chest-mounted transmitters and head-worn receivers enables relative time-of-flight measurements with sub-centimeter precision.
A four-stage signal processing pipeline combines channel impulse response (CIR) interpolation, template matching, phase refinement, and IMU-based ambiguity resolution. It reduces the typical 10 cm UWB ranging error to a median of 5 mm and reaches a median orientation error of 6.5°, validated against Vicon motion capture. HeadSense gives AI wearables an infrastructure-free, body-centric orientation signal without cameras or external tracking.
The work began as HeadTrack, my M.S. thesis at UIUC (2019), and was published as HeadSense at ACM HumanSys '26, co-located with MobiCom 2026.