[C28] A Sub-Milliwatt and Sub-Millisecond 3-D GazeEstimator for Ultra Low-Power AR Applications

Abstract

In this paper, we present the first hardware work that implements the 3D model-based gaze estimation. The critical factors of real-time gaze tracking are high accuracy and user-friendliness such as low latency and no run-time calibration. The most common methodology is PCCR (Pupil Center Corneal Reflection), which can be classified into 2D regression, and 3D model-based methods. Existing hardware works are based on 2D regression algorithms. Though 2D methods have a simple computation process, they require multiple run-time calibration steps, and are vulnerable to head motions. On the other hand, the 3D model-based method can maintain better accuracy than the 2D method with out run-time calibration steps, and robust to head motions. Therefore, we introduce the first 3D model-based gaze estimator with power and area efficiency, while maintaining the accuracy(0.9°).