Car detection using multi-feature selection for varying poses

This paper presents a novel method of car detection by using the Adaboost algorithm, which is enhanced by the quadratic programming for feature extraction. In this paper, car is divided into many relevant features through their appearances in training samples such as wheel and window. We crop features of object in training images and utilize them for the Adaboost training. The results of the Adaboost training are many sets of weak classifiers corresponding to the relevant features. The quadratic programming is applied to set up the priority order of weak classifiers when they are combined together by their relevant positions for detection. In other words, we utilize the Adaboost as a kernel function for generating the stronger classifier, which is a linear combination of weak classifiers selected by the quadratic programming. The proposed method can provide a high accuracy of object detection by using a few hundred samples for training the Adaboost.