Empirical Study of Boosted Weak Classifier in Object Detection Problem

In this paper, we study the use of boosted weak classifiers selected with AdaBoost algorithm in object detection. Our work is motivated by the good performance of AdaBoost in selecting discriminative features and the effectiveness of Classification and Regression Tree (CART) compared with other classification methods. First, we study the cascaded structure of the boosted weak classifier detectors where each boosted weak classifier in the CART is a stump tree that contributes a vote for obtaining the final decision. We then use the same boosted weak classifiersand train them using Multilayer Perceptron (MLP) and Support Vector Machine (SVM) to construct the detectors. We then compare these selected weak classifiers in the cascaded CART detector against the MLP and SVM detector. In this paper, we conduct experiments on the face detection problem and pedestrian detection problem. We show that the cascaded boosted weak classifiers used in MLP and SVM detector outperform the cascaded CART detector.