Natural Grasp Intention Recognition Based on Gaze in Human-Robot Interaction.

Area of Science:

• Neuroscience
• Human-Computer Interaction
• Robotics

Background:

• Neuroscience research links vision and intention, but gaze data features for intention recognition are lacking.
• Existing gaze-based methods rely on long-term fixation, leading to insufficient accuracy.
• There's a need for advanced features to improve grasp intention recognition from gaze data.

Purpose of the Study:

• To suppress noise in human gaze data.
• To extract novel features for accurate grasp intention recognition.
• To address limitations of existing gaze-based intention recognition methods.

Main Methods:

• Gaze movement evaluation experiments were conducted.
• A target-attracted gaze movement model (TAGMM) was proposed for quantitative gaze description.
• A Kalman filter (KF) was used for noise reduction.
• Four novel features (gaze point dispersion, movement, head movement, gaze-object distance) were extracted.
• Intention recognition experiments were performed using various classifiers.

Main Results:

• Proposed features showed significant differences across intentions.
• The TAGMM and proposed features achieved accuracy improvements of 44.26% (within-subject) and 30.67% (cross-subject) over fixation-based methods.
• Intention recognition time was reduced to 34.87 ms, compared to about 1 s for the fixation-based method.

Conclusions:

• The study introduces a novel TAGMM and practical features for recognizing grasp intentions.
• Experimental results confirm the effectiveness of the proposed approach.
• This work advances gaze-based human-robot interaction by enabling better intention recognition.