TraceRing: Touchpad-like Pointing with a Single IMU Ring through Personalized Learning

Paper Title

TraceRing: Touchpad-like Pointing with a Single IMU Ring through Personalized Learning

Publication Info

• Topic area: Wearable technology for human-computer interaction
• Keywords: IMU, touchpad, wearable interaction, personalized learning, cursor control, machine learning, AR/VR, gesture input, Fitts’ Law, usability evaluation

Background and Problem

• Problem / challenge: Achieving precise 2D pointing with a single IMU ring is challenging due to incomplete motion data and significant user variability. Prior single-IMU solutions often underperform compared to multi-sensor or multi-ring setups.
• Significance: A portable, accurate, and ubiquitous pointing solution could enhance interaction with AR/VR devices, large displays, and mobile computing platforms, addressing limitations of current input methods.
• Motivation and related work: Previous approaches using IMU-based rings or other wearable devices often rely on additional sensors or calibration, compromising portability. Single-IMU systems lack accuracy due to physiological variability and motion mapping challenges. This paper builds on these findings and proposes a personalized learning framework to overcome these limitations.

Solution

• Proposed approach: TraceRing, a single-IMU ring system that employs a personalized learning framework to enable touchpad-like pointing.
• Novelty:
  1. Introduction of a multi-task personalization framework combining supervised and contrastive learning for user-specific adaptation without calibration.
  2. Development of a Mixture-of-Experts (MoE) architecture to dynamically select suitable models based on user embeddings.
  3. Demonstration of real-time usability and performance gains over state-of-the-art IMU-based systems.
• Procedure and key techniques:
  • Data preprocessing: Gravity compensation to eliminate irrelevant accelerometer readings.
  • Task Encoder: TCN-LSTM architecture for capturing task-related temporal patterns.
  • User Encoder: BiLSTM with temporal attention for extracting user-specific features.
  • MoE Module: Multiple MLP experts dynamically weighted by user embeddings for velocity prediction.
  • Training: Multi-component loss combining velocity prediction, contrastive learning, and prototype-based clustering.
  • Real-time inference: Event-driven touch detection, CD gain application, and efficient user embedding updates.

Results

• Concrete findings:
  • Velocity prediction error reduced by 33.9% compared to MouseRing (1.48 cm²/s² vs. 2.24 cm²/s²).
  • Real-time task completion time improved significantly over AirMouse (2.26s vs. 3.01s).
  • User embedding effectively distinguishes unseen users and generalizes well.
• Advantage over baselines: TraceRing outperformed MouseRing and AirMouse in velocity prediction and task completion efficiency, while maintaining portability and usability.
• Experiments / evaluation:
  • Dataset: 50 participants, multimodal data collection including IMU, motion capture, and pressure pad.
  • Fitts’ Law usability test: Compared TraceRing to touchpad, mouse, and AirMouse across varying task difficulties.
  • Surface tests: Evaluated performance on different materials (wood, acrylic, rubber, plastic, thigh).
  • Subjective feedback: High user acceptance (87.5% willing to adopt TraceRing) and envisioned diverse application scenarios.
• Limitations and future work:
  • Precision issues for fine-grained control and small targets.
  • Latency due to BLE communication and system delays.
  • Limited dataset coverage for soft or non-flat surfaces.
  • Future improvements include adaptive DPI adjustment, enhanced click detection, and expanded dataset for cross-surface generalization.

Summary

TraceRing introduces a single-IMU ring system that achieves touchpad-like pointing through personalized learning, addressing limitations of prior single-IMU approaches. It reduces velocity prediction error by 33.9% and significantly improves task completion times compared to AirMouse. Usability studies confirm its portability, ease of learning, and high user acceptance, with applications spanning AR/VR, large displays, and mobile computing. While challenges remain in precision and latency, TraceRing demonstrates the feasibility of lightweight, personalized wearable interaction and sets the stage for future advancements in sensor-based machine learning.