The Efficiency Tradeoffs of Gaze Tracking

Abstract

Gaze estimation is a core component of assistive tech- nology and accessibility systems. Existing gaze estimation methods are commonly divided into three major categories: 2D mapping- based, 3D model-based, and appearance-based approaches [1]. While 3D geometric systems can achieve high precision, they often require specialized calibration procedures, complex modeling, and dedicated hardware. Appearance-based deep learning methods are highly effective under natural conditions, but they typically require large annotated datasets and substantial computational resources [3]. In assistive settings, additional challenges such as unintended activation, increased cognitive load, and deployment constraints further complicate system design [2]. This paper presents a hybrid gaze tracking framework for real-time operation on RGB cameras. The proposed system combines geometric iris modeling, CLAHE-based contrast normalization, radial gradient validation, Kalman filtering, and blink-triggered facial landmark analysis. Experimental evaluation shows a mean performance of 75.36 FPS with a 68.38% reduction in positional variance, demonstrating a computationally efficient and accessible alternative for assistive gaze interaction.