Algorithmic gaze annotation for mobile eye-tracking.

Mobile eye-tracking is increasingly used to study human behavior in situ; however, the analysis of the footage is typically performed manually and therefore is slow and laborious. The aim of this study was to examine the extent to which the footage obtained using mobile eye-tracking could be annotated automatically using computer vision algorithms. We developed an open-source Python package that combined two computer vision algorithms to automatically annotate human-body-related areas of interest when two participants interacted with each other. To validate the algorithm, three experienced human raters coded the gaze direction with respect to one of seven a priori defined areas of interest during the task. To test the reliability of the algorithm, the agreement between the human raters was compared with the results obtained from the algorithm. A total of 1,188 frames from 13 trials were compared, with the results revealing substantial agreement between the algorithm and human raters (Krippendorff's alpha = 0.61). The algorithm strictly annotated whether gaze was within or outside of the specified areas of interest, whereas human raters seemed to apply a tolerance when gaze was lying slightly outside the areas of interest. In sum, the computer algorithmic approach appears to provide a valid means of automatically annotating mobile eye-tracking footage in highly dynamic contexts. The possibility of automatically annotating eye-tracking footage of human interactions allows for automatic assessment of visual attention, gaze, and intentions across sectors such as educational settings, pedestrian navigation, and sport.