An accurate eye localization approach for smart embedded system

Abstract

Eye localization is a vital procedure in many applications, such as face recognition and gaze tracking, and can further facilitate related procedures. Although many works have been devoted to localizing eyes in frontal facial images, most approaches cannot work effectively and efficiently in smart embedded systems (e.g., the vehicle system). In this paper, we propose an accurate eye localization approach for smart embedded systems. An illumination normalization procedure with the perception based model is utilized to remove the illumination effects of facial images. Then the integral projection method is employed to localize the candidate positions of eyes. The support vector machine (SVM) classifiers are trained with the spacial and intensity information to verify these candidates rapidly with compact 3-dimensional features. Based on the output of SVMs, the two candidates with top scores are determined as the final accurate eye positions. Extensive experiments on the extended Yale B, AR and ORL face datasets demonstrate that the proposed approach achieves good accuracy and fast computation results for localizing eyes.