Abstract: This paper introduces an advanced gaze-driven cursor control system, exemplifying subject expert excellence in human-computer interaction (HCI) and assistive technology, enabling seamless hands-free computer operation for motor-impaired users through precise eye movements and blinks captured via standard webcam footage. The hybrid CNN-LSTM deep learning architecture at its core employs convolutional layers for high-fidelity extraction of spatial eye features—including pupil centroid, iris boundaries, and geometric landmarks—from real-time video frames, coupled with LSTM recurrent units that adeptly model temporal dependencies to forecast gaze trajectories with sub-pixel smoothness and jitter below 1 pixel variance, while blink detection attains surgical precision (>98% accuracy across diverse head poses) via Eye Aspect Ratio (EAR) derived from eyelid contours and an optimized Support Vector Machine (SVM) robust to occlusions and micro-expressions. User-centric calibration further refines gaze-to-screen homographic mapping through adaptive gain constants, dead zones suppressing physiological noise such as saccades, and dynamic sensitivity regions yielding sub-degree estimation errors (30 minutes) confirming blink precision >97% and pointer control F1-scores >0.95—unequivocally demonstrating consumer-grade hardware's parity with commercial eye-tracking systems in affordability, accessibility, and production viability.

Keywords: gaze-driven cursor, CNN-LSTM, eye tracking, blink detection, assistive technology, webcam-based HCI.

Downloads: | DOI: 10.17148/IARJSET.2026.13135