Hierarchical approach for human eye tracking
Zhihong Zeng, Songde Ma · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2003
This paper presents a hierarchical approach for human eye tracking. The tracking process is accomplished in three steps: head outline tracking, head pose recovery and eye localization. First, active particle filtering is presented to track a head outline. Compared with the traditional particle filtering, the active particle filtering improves the efficiency of every particle and accordingly reduces the number of required particles. In order to recover the head pose, only one view of the head is used to build the shape/texture template including multiple head features, and then the method "combination of views" (Ullman 1996) is applied to model changes of shapes and textures of the head features during head rotation. Based on the result of head pose recovery, the eyes can be localized efficiently in a restricted search space. The algorithm successfully tracks eyes of human heads that undergo translation, scaling and wide-range rotation in the office scene. The non-optimized implementation runs at about 3 frames per second on Pentium III 450MHz PC.