Ways of improving the precision of eye tracking data: Controlling the influence of dirt and dust on pupil detection

Potential of a laser pointer contact lens to improve the reliability of video-based eye-trackers in indoor and outdoor conditions

Many video-based eye trackers rely on detecting and tracking ocular features, a task that can be negatively affected by a number of individual or environmental factors. In this context, the aim of this study was to practically evaluate how the use of a scleral contact lens with two integrated nearinfrared lasers (denoted CLP) could improve the tracking robustness in difficult lighting conditions, particularly outdoor ones. We assessed the ability of the CLP (on a model eye) to detect the lasers and to deduce a gaze position with an accuracy better than 1° under four lighting conditions (1 lx, 250 lx, 50 klux and alternating 1lx /250 lx) on an artificial eye. These results were compared to the ability of a commercial eye tracker (Pupil Core) to detect the pupil on human eyes with a confidence score equal to or greater than 0.9. CLP provided good results in all conditions (tracking accuracy and detection rates). In comparison, the Pupil Core performed well in all indoor conditions (99% detection) but failed in outdoor conditions (9.85% detection). In conclusion, the CLP presents strong potential to improve the reliability of video-based eyetrackers in outdoor conditions by providing easy trackable feature.

Ground Truth Data Generator for Eye Location on Infrared Driver Recordings

Labeling is a very costly and time consuming process that aims to generate datasets for training neural networks in several functionalities and projects. In the automotive field of driver monitoring it has a huge impact, where much of the budget is used for image labeling. This paper presents an algorithm that will be used for generating ground truth data for 2D eye location in infrared images of drivers. The algorithm is implemented with many detection restrictions, which makes it very accurate but not necessarily very constant. The resulting dataset shall not be modified by any human factor and will be used to train neural networks, which we expect to have a very good accuracy and a much better consistency for eye detection than the initial algorithm. This paper proves that we can automatically generate very good quality ground truth data for training neural networks, which is still an open topic in the automotive industry.