Online-EYE: Multimodal Implicit Eye Tracking Calibration for XR

Research Background and Issues

• Identified Challenges and Issues:

  • Eye tracking in extended reality (XR) requires personalized calibration. Traditional explicit calibration methods demand users to recalibrate each time they use the device, which is a tedious task prone to failure due to user inaccuracy.
  • The calibration process interrupts user tasks and affects the overall experience. Additionally, actions like device slippage can cause calibration drift.
  • While generalized models exist for hand tracking, explicit calibration is still required for eye tracking. This raises the question of whether generalized "universal models" and implicit calibration can reduce the need for explicit calibration.

• Significance of the Study:

  • Eye tracking has become a core input technology in XR interactions, such as combining gaze with gestures.
  • Achieving seamless calibration will lower the barrier to device use, optimize user experience, and be applicable in practical scenarios like shared devices.

• Motivation and Related Work:

  • Current work on explicit and implicit calibration faces several limitations: explicit calibration interrupts tasks and its accuracy depends on user cooperation, while implicit methods require further improvement to enhance real-time performance and accuracy.
  • Research on eye-hand coordination and interaction shows that users typically gaze at targets when triggering clicks. This finding provides a new possibility for implicit calibration.

Solution

• Proposed Method or Solution:

  • "Online-EYE" is proposed, a multimodal implicit eye-tracking calibration method based on users' target-clicking behavior in interactive interfaces.
  • The method analyzes the relationship between users' gaze data and the positions of controller clicks, using machine learning techniques to optimize the calibration matrix in real time.
  • It assumes that users accurately gaze at targets when clicking on them, and this assumption was validated across various UI (graphical interface) environments.

• Innovative Aspects of the Method:

  • Compared to traditional explicit calibration, Online-EYE achieves real-time, background implicit calibration without interrupting user tasks.
  • Online calibration does not require large-scale training data and achieves accuracy close to traditional explicit calibration with only 8-9 UI clicks.
  • It provides a seamless transition from a default generalized model to personalized user calibration.

• Implementation Steps and Technical Details:

  1. Validation of Basic Assumptions:
     • Experiments were designed to verify whether users gaze at targets when clicking on them (assumption confirmed).
  2. Implicit Calibration Based on UI Elements:
     • A generalized model (calibration data from the previous user) was simulated. Through various UI tasks, deviations in target clicks were recorded, and the calibration matrix was optimized using the recursive least squares algorithm.
  3. Real-Time Calibration Algorithm:
     • The recursive least squares (RLS) algorithm was used, dynamically adjusting weights via a "forgetting factor" to update the calibration matrix in real time.
  4. Validation in Application Scenarios:
     • The practical application of implicit calibration was evaluated in the "Space Invaders game" and "user survey" scenarios.

Research Outcomes

• Specific Results:

  • Implicit calibration significantly improved the eye-tracking accuracy of the initial generalized model across multiple experimental tasks, achieving performance comparable to explicit calibration in certain scenarios.
  • In experiments, Online-EYE enabled precise gaze selection functionality with an average of 8-9 clicks.

• Comparison with Existing Solutions:

  • Compared to explicit calibration, implicit calibration achieves dynamic calibration without interrupting user tasks and performs stably during frequent user-system interactions.
  • Compared to other implicit calibration techniques, Online-EYE is applicable to various UI interaction types (buttons, drag-and-drop, sliders, etc.) and achieves accuracy comparable to explicit calibration.

• Experimental or Evaluation Results:

  • In evaluation methods, the average error of implicit calibration (1.94°) was significantly better than the uncalibrated generalized model (3.08°). Explicit calibration performed better (1.19°) but showed significant advantages mainly in peripheral areas of the visual field.
  • Working distance (0.5 meters and 2 meters) did not significantly affect the accuracy of implicit calibration.
  • User experience validation in two application scenarios demonstrated that implicit calibration achieved a level of smoothness unmatched by manual calibration.

• Limitations and Future Directions:

  • The calibration range is limited by the field of view covered by UI elements, leading to inaccuracies in boundary areas. Future strategies could involve designing UI elements to cover a broader field of view.
  • Users may exhibit "compensatory behavior" to adjust their gaze when calibration accuracy is insufficient, affecting the calibration algorithm's precision. Future research could integrate smarter cursor adjustment methods or allow users to manually switch input modes.
  • The current study focuses on controller operations in VR. Future work should verify the applicability of this implicit calibration model to other input methods, such as gesture control.

In summary, the Online-EYE method significantly reduces the complexity and time cost of explicit calibration, opening new use cases and potential for eye tracking in XR. It lays a foundation for advancements in multimodal interaction technologies.