ISWC '25: ACM International Symposium on Wearable Computers • 2025

From Neck to Head: Bio-Impedance Sensing for Head Pose Estimation

Mengxi Liu (German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany), Lala Shakti Swarup Ray (German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany), Sizhen Bian (German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany), Ko Watanabe (German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany), Ankur Bhatt (German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany), Joanna Sorysz (German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany), Russel Torah (University of Southampton, Southampton, United Kingdom), Bo Zhou (German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany), Paul Lukowicz (German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany)

Bio-impedance SensingPose EstimationTextile Electrode

Highlights

  • Necklace-style wearable with five soft, dry, reusable electrodes measures neck bio-impedance (magnitude and phase).
  • Imp2Head transformer maps impedance sequences to SMPL-X head, neck, and jaw rotations with biomechanical constraints.
  • Leave-one-person-out study (N=7) achieves MPVE 5.9 mm vs pseudo ground truth and an estimated corrected MPJPE of 25.9 mm vs MoCap.
  • Simplified hardware using AD5941 AFE and ESP32-S2 with Bluetooth; no EIT image reconstruction required.

Abstract

We present NeckSense, a wearable system for head pose tracking that uses multi-channel bio-impedance sensing with soft, dry electrodes in a lightweight necklace. NeckSense captures impedance changes around the neck modulated by head rotations and subtle muscle activations. We propose Imp2Head, a deep learning framework that integrates anatomical priors, including joint constraints and natural head rotation ranges, into the loss design to map impedance features to 3D head, neck, and jaw rotations. Using a simplified five-electrode configuration (one common stimulation and four measurement channels) and frequency-domain impedance features, the system avoids tomographic reconstruction and reduces hardware and computational complexity. Evaluated on 7 participants with a leave-one-person-out protocol and camera-based pose estimation as pseudo ground truth, NeckSense achieves a mean per-vertex error of 5.9 mm versus the vision model and an estimated corrected MPJPE of 25.9 mm relative to motion capture, demonstrating performance comparable to state-of-the-art vision methods without requiring line-of-sight.

Click to read Full Paper.

Want to stay up to date with all the information?

Sign up for our newsletter!