Realtime and Accurate 3D Eye Gaze Capture with DCNN-based Iris and Pupil Segmentation.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Machine Learning

Background:

• Accurate 3D eye gaze tracking is crucial for various applications, including virtual reality, augmented reality, and assistive technologies.
• Existing methods often require specialized hardware or multiple cameras, limiting their accessibility and real-world applicability.
• Robustness to variations in lighting, facial expressions, and eye conditions like blinking remains a challenge.

Purpose of the Study:

• To develop a real-time, accurate, and robust 3D eye gaze tracking system using a single monocular RGB camera.
• To leverage deep learning for automatic extraction of eye features and robust gaze estimation.
• To enhance system performance by incorporating eye-blink detection.

Main Methods:

• A deep convolutional neural network (DCNN), combining Unet and Squeezenet architectures, was trained for pixel classification to extract iris and pupil regions.
• The Maximum A Posteriori (MAP) framework was employed for sequential tracking of the 3D eye gaze state.
• An extension of the DCNN was developed for eye-close detection to improve tracker robustness during blinks.

Main Results:

• The system achieves real-time performance on both desktop PCs and smartphones.
• Demonstrated robustness and accuracy across diverse genders, races, lighting conditions, poses, facial shapes, and expressions.
• Outperformed previous state-of-the-art methods in 3D eye tracking using a single RGB camera, as evidenced by comparisons.

Conclusions:

• The proposed DCNN-based method offers a significant advancement in single-camera 3D eye gaze tracking.
• The system's real-time capability and robustness make it suitable for widespread practical applications.
• The integration of eye-blink detection further enhances the reliability of the eye gaze tracker.