A Revised Eye-Localization Algorithm Based on Random Forest

In this paper, we present an enhanced Random Forest(RF) model for precise eye localization. To extend Random Forest, we 1) propose the randomized trees with adaptive gradient boosting for a more accurate eye localization 2) introduce a series of standard samples with random perturbation for the robustness to changes in illumination and head pose and eye rotation. Performance results of our methods showed that it can obtain an accuracy of 92 percent on BioID database and gives a frame processing time of less than 1ms because of the low computation cost. Experimental results on the challenging BioID database show that our model can locate eyes accurately and efficiently under a broad range of uncontrolled variations involving lightings, camera qualities, occlusions, etc.