.png){kind=icon}
.png){kind=icon}
Return
Lightweight webcam-based eye tracking system for large display screens
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-026-21208-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The existing webcam-based eye-tracking methods are often inaccurate when applied to large-screen unmanned automatic kiosks, due to significant camera-to-user distances, head pose variability, and time-consuming calibration procedures, limiting their suitability for public kiosk applications. This research aims to develop and evaluate a novel, accurate webcam-based eye-tracking system specifically designed for interaction with large-screen kiosks, overcoming the challenge of camera distance. We propose a spatial attention–based deep learning feature extractor to obtain high-fidelity 3D facial mesh and iris landmarks under varied head poses and distances, coupled with lightweight machine-learning regression models for screen coordinate prediction. In addition, we proposed a novel smooth-moving calibration scheme with adjustable speed to reduce calibration time and improve user engagement. The system was tested on a custom-built kiosk featuring a 32-inch display (1080 $$\times$$ 1920 pixels). Experiments compare four regression models, including Linear Regression with Stochastic Gradient Descent (SGD) and Ridge Regression, each tested with four distinct user calibration methods. The smooth-moving calibration point proved to be the most effective calibration method. The SGD model achieved the highest accuracy at 96%, with an average pixel error of 73.42px on the x-axis and 140.99px on the y-axis. The Ridge Regression model also performed well, obtaining 84% accuracy. Notably, when the calibration time for the Ridge Regression model was reduced from 80 to 40 seconds, its average pixel error improved significantly to 48px on the x-axis and 87px on the y-axis. This study demonstrates the successful implementation of a high-accuracy eye-tracking system for large-screen kiosks using only a standard webcam. The proposed method, particularly the combination of Ridge Regression with a shortened calibration time, presents a robust and efficient solution that is practical for real-world deployment in interactive public systems.
Keywords:
Eye tracking
Facial landmark extraction
Human computer interface
Screen calibration
Screen coordinate prediction
Journal
IF:
3
Papers:
1.9W
Citations:
3.2W