静态图像中基于多分类器的人体检测技术研究

Human Detection Based on Multi-classifiers in Static Image

首先介绍Haar特征,然后介绍用于分类器训练的Adaboost算法,该方法训练的级联分类器用于人体检测时虽然具有很高的检测率,但虚警率较高.为了保持检测率,降低虚警率,在原有分类器的基础上再训练两个分类器,一个是利用头肩样本训练的分类器,另一个是利用腿部样本训练的分类器.实验证明:该方法设计的分类器在保持较高的检测率的同时其虚警率比原方法设计的分类器降低一个数量级.

In this paper, we introduce Haar feature firstly, and then interpret Adaboost arithmetic which is used in training classifier. The false positive rate of this arithmetic is too high although the detecting rate is so high. In order to maintain detecting rate and reduce false positive rate, we train another two local classifiers： one is head- shoulder classifier, the other is legs classifier. The experiment shows the new method not only have high detecting rate but also can reduce false positive rate.