Adaptive Eye-Controlled Mouse using Deep Belief and Q-Learning Networks

With increasing use of traditional input devices such as keyboards and mice, people with motor impairments would face many more difficulties in accessing computers. The proposed system of an Eye-Controlled Mouse Pointer presents a hands-free, vision-based solution for intelligent humancomputer interaction. The system captures real-time eye movements using a standard webcam; Python's OpenCV and Dlib are used for facial landmark detection and tracking of the pupil. The visual features extracted are further enhanced through the use of a Deep Belief Network to ensure gaze and blink recognition that is robust across changing lighting conditions and head orientation. A reinforcement model based on a Deep Q-Network is trained which, through reward adaptation, dynamically learns the most effective cursor movements and click actions. Calibration and integration testing of the hybrid model are facilitated by MATLAB. The novelty in its contribution, this work fuses DQN and DBN to realize adaptive, highly accurate, low-latency mouse control that allows an accessible, cost-efficient alternative to physical input devices.