EyeCRNet: eye movement segmentation and pupil ellipse fitting for unconstrained outdoor environments

This paper aims to segment eye movements in videos captured in outdoor environments and extract the parameters of the ellipse fitted to the pupil, thereby improving the accuracy of gaze tracking in natural settings. Current eye-tracking methods face challenges in complex outdoor environments due to lighting variations, head movements, eyelid occlusions, and changes in camera angles, leading to reduced accuracy in eye segmentation and pupil parameter extraction. To address these issues, I propose a framework, EyeCRNet, designed for pupil fitting in outdoor eye movement analysis. This framework includes two main components: eye segmentation and pupil ellipse fitting, to enhance the robustness of pupil segmentation in outdoor conditions. First, the EyeCRNet model is trained on the OpenEDS2019 dataset to ensure accurate eye segmentation. Then, videos from the LPW dataset are segmented into 500 frames, and manual segmentation of the pupil, iris, and sclera is performed on these eye movement video frames. EyeCRNet utilizes the video dataset and employs an attention mechanism to strengthen the association between the pupil and corneal reflection points (CR). Through multi-task learning, the model effectively improves the detection of the pupil and corneal reflection points (CR) in natural environments, enhancing prediction stability and robustness. Additionally, it accurately calculates the geometric parameters of the fitted pupil ellipse, including center position, major and minor axes, and rotation angle. Experimental results demonstrate that, the EyeCRNet network effectively adapts to lighting and dynamic eye changes.