Multi-User and Multi-View Human Eyes' Detection and Tracking

Abstract

This paper presents a framework on multi-user and multi-view human eyes' detection and tracking. First, it uses fives kinds of AdaBoost face detectors with four different sizes at each area of image to detect faces in turn. Then, to locate eyes' positions, four kinds of AdaBoost eye detectors are used and if the eye-detection above fails, the prior knowledge of human organs' positions in anatomy is applied as a spare method. Next, it uses the unscented filter to predict the targets' next possible positions. Finally, the detection method above is used to detect the third frame and amend the relative forecasting. And repeat above cycle until tracking over. This framework is robust to subject's eyes' blinking, closing, wearing glasses and partly sheltering in multi-face and multi-view to a certain extent for the optimized structure performance and reasonable selected features. And because of the nonlinear filtering, it can track targets in curves with changing speeds. It mainly fits most usual vertical head scenes in monitoring environment.