RadEye: Tracking Eye Motion Using FMCW Radar

Research Background and Issues

• Identified Problems or Challenges: Existing contactless eye-tracking technologies (e.g., camera-based, acoustic, or millimeter-wave radar methods) face the following issues:

  1. Cameras, while accurate and easy to use, may raise privacy concerns and perform poorly under low-light conditions.
  2. Acoustic methods are limited by propagation distance and are generally suitable for short-range blink detection.
  3. Millimeter-wave radar can achieve millimeter-level motion detection but is constrained in detection range and primarily focuses on blink detection rather than eye-tracking.

• Significance: Eye-tracking holds significant value across various scenarios, including human-computer interaction (HCI), virtual reality, psychological research, market analysis, and early disease detection. In certain cases, such as assisting patients with amyotrophic lateral sclerosis (ALS) in communication, privacy protection and non-invasive solutions are particularly critical.

• Research Motivation and Related Work: The motivation of this study is to develop a long-range contactless eye-tracking system that overcomes the limitations of existing methods, ensuring privacy and reliability, especially under poor lighting conditions. The authors aim to address the research gap in long-range eye-tracking using RF signals.

Solution

• Method or Solution: The authors propose a system called RadEye, which leverages sub-6GHz FMCW radar combined with deep neural networks (DNN) to achieve millimeter-level eye movement detection and long-range tracking.

  1. Hardware: A custom-designed 5GHz FMCW radar was developed to achieve the required detection resolution and range.
  2. Software: A transformer-based DNN structure supervised by camera data was employed to further enhance detection accuracy.

• Innovations:

  1. RadEye is the first system to estimate eye movement angles at long distances using RF signals.
  2. By combining hardware design (e.g., custom antennas) and software optimization (e.g., camera-guided DNN), the system effectively addresses challenges in fine eye movement detection and interference suppression.
  3. Compared to existing technologies, RadEye significantly expands the detection range (from less than 1 meter to over 5 meters) and achieves high-accuracy eye movement direction tracking.

• Implementation Steps and Key Techniques:

  1. Signal Processing:
     • FMCW radar sends frequency-modulated signals and receives reflected signals from targets to extract eye movement features.
     • Range FFT is used to distinguish signals reflected by different objects, and sliding window detection captures dynamic changes in user eye blinks.
  2. Feature Extraction:
     • Radar signal amplitude and phase changes caused by eye muscle and eyelid movements are modeled and separated using specific features (e.g., amplitude variation rate and motion curvature in the complex domain).
  3. Deep Learning Model:
     • A transformer-based DNN structure encodes input signals to predict eye movement angles.
     • Eye movement angles generated by cameras are used as supervised labels to train the model, ensuring its generalization capability.

Research Outcomes

• Specific Results:

  1. RadEye achieves high accuracy (90% direction detection accuracy) in contactless eye-tracking and supports detection distances up to 5 meters.
  2. When users provide eye movements in different directions, the system's parameter estimation error is only 24 degrees (azimuth) and 21 degrees (elevation).

• Advantages:

  1. Privacy Protection: RF signals cannot directly resolve personal facial features, significantly reducing privacy risks compared to cameras.
  2. Extended Adaptability: The system operates effectively in low-light environments and supports zero-shot detection for more users and scenarios.

• Experiments or Evaluations:

  1. Experimental Setup: RadEye's performance was evaluated at three distances (3 meters, 4 meters, and 5 meters) and under various background environments.
  2. Interference Adaptability:
     • RadEye can handle environmental interference (e.g., nearby moving objects).
     • It maintains a certain level of accuracy even during user movements (e.g., head, mouth, or limb activities).
  3. Zero-Shot Generalization:
     • RadEye demonstrates excellent adaptability to untrained users and new scenarios.

• Limitations and Future Directions:

  1. Limitations:
     • Users must keep their heads still, as head and mouth movements may interfere with detection.
     • The current prototype is bulky and unsuitable for certain scenarios (e.g., wheelchair-mounted applications).
     • User eye movement inputs are limited to four directions, potentially causing fatigue.
  2. Future Directions:
     • Increase radar bandwidth to separate mouth movement interference.
     • Optimize device size for more lightweight and compact designs.
     • Enhance model resolution to support continuous eye movement direction tracking, improving input efficiency and reducing user fatigue.

This paper demonstrates the potential of RF sensing technology in remote eye-tracking applications and provides a solid foundation for future advancements in related technologies.