Exploration of webcam-based eye tracking by comparing gaze behaviour of Dutch Sign Language users during interpretation of sign language

Abstract

The present study investigates properties of webcam-based eye tracking for online experiments using the WebGazer JavaScript library implemented in Gorilla Online Experiment Builder. Participants (n=15) completed a calibration sequence to optimize the eye tracking quality of their webcam, after which they were they were tasked to attentively watch 9 video’s in which a story was signed in Dutch Sign Language. Between video 4 and 5 calibration was repeated. The main research question is: what are the accuracy and precision of a webcam-based eye tracker for online experiments? The accuracies (reported in degrees of visual angle) of the webcams averaged 3.65° (SD = 0.77°) and the precisions (reported in inter-sample distance root mean square (RMS)) averaged 0.89° RMS (SD = 0.34° RMS). Furthermore, the data was of sufficient quality to filter out typical properties of gaze behaviour such as fixations, smooth pursuit eye movements and blinks. These results show promise that online experiments with a webcam-based eye tracker can be a viable alternative to lab-based experiments with classic infrared eye trackers in cases where very accurate data quality is not required. The eye tracking data drifted during the experiment causing a loss in accuracy. Attempts to improve accuracy by post-hoc compensation of the drift were unsuccessful. The author proposes further research attempts to determine the origin of this drift and possible remedies.