Research on Eye-Tracking Control Methods Based on an Improved YOLOv11 Model

Eye-tracking technology has gained traction in the field of medical rehabilitation due to its non-invasive and intuitive nature. However, current eye-tracking methods based on object detection technology suffer from insufficient accuracy in detecting the eye socket and iris, as well as inaccuracies in determining eye movement direction. To address this, this study improved the YOLOv11 model using the EFFM and ORC modules, resulting in a 1.7% and 9.9% increase in recognition accuracy for the eye socket and iris, respectively, and a 5.5% and 44% increase in recall rate, respectively. A method combining frame voting mechanisms with eye movement area discrimination was proposed for eye movement direction discrimination, achieving average accuracy rates of 95.3%, 92.8%, and 94.8% for iris fixation, left, and right directions, respectively. The discrimination results of multiple eye movement images were mapped to a binary value, and eye movement encoding was used to obtain control commands that align with the robotic arm. The average matching degree of eye movement encoding ranged from 93.4% to 96.8%. An experimental platform was established, and the average completion rates for three object-grabbing tasks controlled by eye movements were 98%, 78%, and 96%, respectively.