Classifying Head Movements to Separate Head-Gaze and Head Gestures as Distinct Modes of Input

Title

Classifying Head Movements to Separate Head-Gaze and Head Gestures as Distinct Modes of Input

Document Information

• Domain: Human-Computer Interaction, Machine Learning, Virtual Reality Interaction
• Keywords: Head Gestures, Eye Tracking, Virtual Reality, Eye-head Coordination, Computational Interaction, Machine Learning, XGBoost

Research Background and Problem

• Identified Issues or Challenges:

  • Head movements are often regarded as a single type of input, but there are actually two fundamentally different types: Head-Gaze (head movements accompanied by eye movements) and Head Gestures (head movements independent of eye movements).
  • Current head-tracking interfaces do not distinguish between these two input modes, leading to interference, such as false triggers caused by Head-Gaze or missed interactions due to Head Gestures.
  • Existing threshold-based classification methods lack accuracy and fail to effectively separate these two types of movements.

• Significance:

  • Distinguishing between Head-Gaze and Head Gestures can leverage the advantages of each movement mode, providing more precise and natural user experiences in interaction design.
  • In scenarios like Virtual Reality (VR) and Augmented Reality (AR), head-based input is increasingly used for touchless interaction, but classification and filtering mechanisms still need improvement.

• Research Motivation and Related Work:

  • Inspired by Sidenmark et al.'s BimodalGaze method, which attempted to separate Head-Gaze and Head Gestures using threshold-based rules but showed limited effectiveness.
  • Machine learning methods have been preliminarily applied to modeling head movement behaviors, but no specific solutions have been proposed for distinguishing Head-Gaze and Head Gestures.

Solution

• Proposed Method or Solution:

  • Developed a machine learning classifier named HeadBoost, which automatically distinguishes Head-Gaze and Head Gestures based on extensive training data.
  • Utilized the XGBoost algorithm, incorporating various features (e.g., morphological, noise, temporal, and correlation features) to significantly improve classification accuracy.

• Innovations:

  • Introduced a clear distinction between gaze-driven and gaze-independent head movements and implemented classification.
  • Unlike previous threshold-based classifiers, HeadBoost captures more complex movement patterns and features through a machine learning model.
  • Designed unique experimental tasks and data collection methods to automatically label classification data, avoiding human errors introduced by subjective manual labeling.

• Implementation Steps and Techniques:

  1. Data Collection:

     • Designed tasks in a virtual reality environment where participants alternated between eye-focused movements (Head-Gaze) and head gestures (Head Gestures).
     • Recorded data such as head position and eye direction, automatically labeling them as Head-Gaze or Head Gestures.
     • Collected over 1.19 million timestamped data samples from 18 participants.

  2. Feature Extraction:

     • Extracted five categories of features, including morphological features (e.g., speed, acceleration), noise features (e.g., standard deviation, dispersion), temporal features (e.g., time since the last saccade), spectral features, and correlation features.
     • Used correlation filtering and other methods to select 81 most relevant features from multiple candidates.

  3. Model Development and Training:

     • Trained the XGBoost model, validating the classifier's performance using both user-independent and user-dependent approaches.
     • Tested different time windows and sampling rates to optimize model performance.

Research Outcomes

• Specific Results:

  • HeadBoost achieved high classification performance in user-independent tests, with an ?1-Score of 0.89, significantly outperforming baseline methods (?1-Score: 0.62).
  • The classifier addressed the baseline method's inability to detect low-speed head gestures and could detect gesture initiation 119ms earlier.

• Advantages Compared to Existing Methods:

  • Significantly improved classification performance.
  • Effectively distinguished slow or fine head gestures without relying on fixed thresholds.
  • Established a general machine learning-based model adaptable to individual user differences.

• Experimental or Evaluation Results:

  • Metrics such as ?1-Score and AUC for the user-independent model outperformed the user-dependent model, demonstrating the method's applicability to general scenarios.
  • Refined the applicable range of head movements: minimum controllable rotation angle of approximately 0.3°, maximum comfortable angle of approximately 40°.

• Limitations and Future Directions:

  1. Limited Experimental Scope:
     • Current experiments required participants to remain seated, allowing only head rotations. Future studies should expand to include torso movements and dynamic environments.
  2. Device Limitations:
     • Data collection relied primarily on VR headsets, and results may be affected by low-precision devices in real-world scenarios.
  3. Real-time Limitations:
     • The current classifier operates at a rate of 30Hz. Future research could optimize the model through dimensionality reduction to improve response speed.
  4. Application Evaluation Insufficiency:
     • Although three example applications were proposed, they have not been extensively validated through user experiments.
  5. Intent Recognition Issues:
     • The classifier cannot directly determine whether head movements have interaction intent. Future work should integrate more complex behavior understanding techniques for optimization.

Conclusion

By designing an efficient machine learning classifier, HeadBoost, this study successfully distinguished Head-Gaze and Head Gestures, providing new possibilities and practical approaches for head-driven human-computer interaction. HeadBoost's innovations enable interaction designs to improve accuracy while offering a more natural experience and avoiding false triggers, making it suitable for emerging applications in AR/VR environments.