부산대학교 도서관

The ability to estimate a person’s gaze has become a vital technology in healthcare and human behavior studies. However, existing commercial eye trackers are expensive and rely on dedicated hardware limiting them to controlled lab settings. RGB camera-based eye trackers face accuracy challenges due to varying illumination, head movement and occlusions. While a good and uniform illumination is key for RGB camera gaze trackers to perform well, infrared-based gaze trackers do not have this issue. Unfortunately, they are not readily available. To address this, we propose CamTratak - a novel framework for gaze tracking, using an off-the-shelf low-resolution RGB camera (Webcam). Our novel framework detects and extracts glint features from eye images and fuses them with eye appearance-based features for gaze tracking, making it robust against varying illumination conditions and suitable for unconstrained settings (without the need for a chin rest). The results generated over the benchmark dataset establish that our framework is robust even in conditions with varying illumination and head pose, and has improved accuracy over the existing benchmark state-of-the-art (Webgazer gaze tracker). The experimental results of our model in real-life use cases such as assisted living/hospital beds and telerobotics, demonstrate the suitability of our model for its widespread adoption in healthcare applications.