Skip to Main Content
A classifier structure is described to construct boosted tree classifier stages for detecting dynamic objects such as human faces. It is based on using simple and weak classifiers based on Haar-like features which are easily computed by integral image representation values (Viola, P. et al., IEEE CVPR, 2001) and putting these weak classifiers into the nodes of a binary tree structure by performing the AdaBoost algorithm. Each classifier stage consists of decision trees containing simple classifiers at their nodes. Each node classifier is constructed using only one simple feature and is divided into sub nodes if the decision tree error on the training set is greater than a predetermined threshold. This division process continues until the error is lower than the threshold and a maximum tree depth is not exceeded. By this approach, it is tried to obtain a faster convergence to a certain classifier performance value.