The article, Hybrid Eye-Tracking on a Smartphone with CNN Feature Extraction and an Infrared 3D Model by Brousseau et al., describes a low-cost, robust and accurate remote eye-tracking system that uses an industrial prototype smartphone with integrated infrared illumination and camera.
In fields such as neurological and neuropsychiatric research, advertisement assessment, pilot training, and automobile safety, multiple research studies have shown benefits from using eye-tracking. The research has demonstrated that specialized eye-tracking systems can be used in many domains, however, the need to purchase expensive, specialized software and hardware to monitor the point-of-gaze (PoG) limits the use of eye-tracking systems by consumers. In this study, an eye-tracking system was integrated into a modern smartphone; therefore, making the eye-tracking system more accessible.
The new hybrid eye-tracking system for smartphones can achieve a gaze-estimation bias of 0.72° in realistic scenarios across 8 subjects. The system also uses machine-learning algorithms to improve the robustness and accuracy of eye-feature estimation. The eye features, such as the location of corneal reflections and the center of the pupil are then used by a gaze-estimation model that is insensitive to relative motion between the subject and the smartphone. A 3D gaze-estimation model is used by the system that allows accurate PoG estimation with free motion of the head and computer. The framework uses Convolutional Neural Networks (CNNs) along with a novel center-of-mass output layer to accurately assess the input eye features. The use of CNNs increases the robustness of the device with considerable variability in the appearance of eye images used in handheld eye trackers.
The hybrid eye-tracking system, which has the advantage of infrared illumination, a 3D gaze-estimation model and a CNN function extractor achieved an accuracy that is 400% higher than the accuracy achieved by previous smartphone eye-tracking systems that use natural light and appearance-based methods to estimate PoG. The improvement in performance comes from the hybrid approach -a machine-learning feature extraction step followed by a geometric 3D model. This improvement may enable mobile applications that require analysis of visual scanning behaviour of subjects that view a limited number of items on a smartphone screen.
Reference
Brousseau B, Rose J, Eizenman M. Hybrid Eye-Tracking on a Smartphone with CNN Feature Extraction and an Infrared 3D Model. Sensors. 2020; 20(2):543. https://doi.org/10.3390/s20020543