EITPose: Wearable and Practical Electrical Impedance Tomography for Continuous Hand Pose Estimation

Title of the Paper

EITPose: Wearable and Practical Electrical Impedance Tomography for Continuous Hand Pose Estimation

Paper Information

• Research Domain: Gesture Recognition and Hand Pose Estimation in Human-Computer Interaction and Mobile Computing
• Keywords: Electrical Impedance Tomography, Hand Pose, Gesture Recognition, Interaction Technique, Wearable Devices, Privacy-preserving Sensors, Machine Learning, Extended Reality

Research Background and Problem

• Problems or Challenges:

  • Hand pose estimation is widely applied in fields such as mixed reality, robotic control, gaming, and sign language recognition. However, current methods primarily rely on cameras, which pose challenges related to privacy, security, and environmental interference, such as lighting and occlusion issues.
  • Many existing portable devices fail to achieve precise continuous gesture tracking or require high power consumption and complex hardware.
  • Existing solutions often lack consistency and accuracy when used across different users or sessions.

• Significance:

  • Developing a privacy-sensitive and lightweight method for digitizing hand movements, especially in a consumer-friendly form factor, remains an unresolved challenge.

• Research Motivation and Related Work:

  • The authors reviewed existing hand sensing technologies, including optical methods (e.g., RGB cameras, depth cameras, and infrared cameras) and non-optical methods (e.g., electrical impedance tomography, EMG, and RF). These methods have limitations in terms of privacy, invasiveness, tracking accuracy, and energy consumption.
  • The medical applications of Electrical Impedance Tomography (EIT) demonstrate its non-invasive and low-cost characteristics, suggesting its potential for extension to hand pose estimation.

Solution

• Proposed Method or Solution:

  • The authors propose EITPose, a wrist-worn device based on EIT for continuous hand pose estimation.
  • EITPose measures internal impedance distribution using 8 electrodes placed around the forearm to infer 3D hand poses.

• Innovations:

  • The device offers a privacy-first, low-power (0.3W), and slim (12mm) design with accuracy comparable to camera-based methods.
  • A waveform inspection algorithm is introduced to maintain signal robustness during user sessions.
  • The device is enhanced with high-pass filtering and automatic initialization features, optimizing skin-electrode contact and reducing noise.
  • A machine learning architecture is designed to ensure consistent performance across different users and sessions.

• Implementation Steps and Key Technologies:

  1. Device Initialization: Automatically optimizes current injection and voltage gain to reduce signal drift caused by skin differences.
  2. Hardware Design:
     • Custom EIT board with high-pass filtering and enhancement modules.
     • 8 stainless steel electrodes with an elastic wristband for stable skin-electrode contact.
  3. Software Optimization:
     • Real-time evaluation of signal validity using a waveform inspection algorithm.
     • Hand pose prediction using a machine learning model based on the ExtraTreesRegressor algorithm.
  4. Data Collection and Analysis:
     • Longitudinal, cross-session, and cross-user studies involving 22 participants.

Research Outcomes

• Specific Results:

  • Achieved the following hand pose estimation accuracies:
    • Within-session mean per-joint position error (MPJPE): 11.06 mm.
    • Cross-session error: 17.81 mm.
    • Cross-user error: 18.91 mm.
  • In gesture recognition tasks, EITPose demonstrated superior cross-user performance compared to existing EIT solutions (e.g., Tomo), with accuracy improvements from 38.8% to 67.3%.

• Comparison with Existing Solutions:

  • Comparable accuracy to camera-based methods like DiscoBand and FingerTrak, with significantly reduced power consumption (from 3.6W to 0.3W).
  • Outperformed acoustic-based methods like BeamBand in cross-user performance for gesture recognition tasks.

• Experimental or Evaluation Results:

  • System performance slightly declines over time but remains more accurate than many related methods.
  • EITPose exhibits robustness in applications with complex gesture variations and numerous poses.

• Limitations and Future Directions:

  • Limitations:
    • Accuracy degradation across sessions and users still requires improvement.
    • The current use of 8 electrodes limits spatial resolution, preventing finer capture of internal forearm structures.
  • Future Directions:
    • Explore the addition of more electrodes to enhance resolution while optimizing sampling rates.
    • Integrate IMU modules to improve sensor adaptability during pose changes.
    • Develop global machine learning models based on larger datasets and refine them with personalized approaches.

Open Source and Social Contribution

• The authors plan to open-source the dataset, processing pipeline, and models to advance related research.
• Open-source repository: GitHub