Investigation of wavelength and intensityeffects in infrared-based eye tracking systemsunder variable lighting and obstructive conditions

Abstract

Eye tracking systems are highly practical and effective tools widely used in research, clinical studies, and as safety systems for vehicle operators. As the name suggests, eye tracking involves monitoring a subject's eye movements to determine gaze points, pupil position, pupil dilation, and blink rates. These systems have a history dating back 200 years, initially relying on human observers tracking participants' reading patterns. Over time, technological advancements in image recording, non-intrusive illumination methods, and image processing algorithms have significantly improved their capabilities. At its core, an eye tracking device consists of a camera focused on the eye and an illumination source. Infrared (IR) light is typically used for illumination to avoid discomfort. The camera is equipped with an IR pass filter to detect infrared wavelengths. The captured images are then processed either in real-time or later to extract the desired features. In video-based systems, heat maps can also be generated by analyzing gaze patterns. There are two main types of eye tracking systems. Fixed or desktop eye trackers integrate the camera and IR illuminators into a stationary unit that captures the user's entire face. These systems are particularly sensitive to head movements and are commonly used with screens or projectors. Mobile eye trackers, resembling glasses, have cameras and illuminators mounted on the frame to minimize motion effects. They include a scene camera that records the environment, enabling gaze mapping in real-world settings such as field marketing studies, sports analysis, and paper-based assessments. Like all electronic devices, these systems are affected by various noise factors. Motion in dynamic environments can cause image jitter, making processing difficult. Poor lighting conditions may reduce feature visibility, while excessive light can create glare and saturation issues. Protective eye-wear like sunglasses or visors can reflect IR light, complicating eye detection. Some research also indicates that eye color may influence tracking accuracy. Currently, eye tracking systems are being implemented in vehicles as driver safety features and are being adapted for pilot monitoring in aircraft cockpits. These environments present the same noise challenges mentioned earlier. While many studies have focused on developing software algorithms to address these issues, this research examines whether captured images are suitable for processing before algorithmic analysis. Specifically, it investigates the effects of eye color, different IR wavelengths, varying light intensities, and protective eyewear on tracking performance. The study was conducted in a laboratory setting designed to simulate a cockpit environment. Three IR light sources (730 nm, 850 nm, and 940 nm) were positioned to illuminate the participant's face, with a frontal camera fixed 80 cm away for image capture. The eye color experiment involved participants with brown and blue eyes tested under day and night conditions at three IR wavelengths with fixed intensity, while they focused on five predefined points. For wavelength and light intensity testing, images were captured under three wavelengths and five current levels in both day and night conditions, with participants focusing on 15 points. The protective eyewear test repeated these conditions with sunglasses. Image processing employed both computer vision algorithms for pupil center and radius detection and convolutional neural networks (CNNs). Ground truth measurements were manually annotated, with errors calculated as the difference between algorithmic and manual results. The findings were organized by wavelength and input current, further divided into total, night, and day conditions. Results showed no significant variation due to eye color. Without sunglasses, computer vision performed best at 850 nm - 0.65 A at night and 850 nm - 1.04 A during daytime, while 730 nm - 0.60 A yielded the best overall results. CNNs achieved optimal performance at 730 nm - 0.12 A across all conditions. With sunglasses, computer vision worked best at 850 nm - 0.65 A at night and 730 nm - 0.60 A during daytime, with 850 nm - 0.17 A performing best overall. CNNs with sunglasses showed best results at 730 nm - 0.12 A at night, 940 nm - 1.5 A during daytime, and 730 nm - 0.12 A overall. In conclusion, the study demonstrates that under basic algorithmic frameworks, in various environmental conditions tracking accuracy varies significantly depending on wavelength and light intensity, . These findings suggest that developing eye tracking systems incorporating multiple wavelengths could improve performance, particularly for automotive and aviation safety applications where reliable operation under varying conditions is crucial. The research highlights the importance of optimizing hardware configurations before software processing to ensure accurate and robust eye tracking performance.