Asymmetric learning for pedestrian detection based on joint local orientation histograms

Abstract

We present a cost-sensitive learning framework for pedestrian detection in still images based on the novel Joint Local Orientation Histograms (JLOH) features and the Asymmetric Gentle AdaBoost. The JLOH features capture the co-occurrence of local histograms and make it possible to classify the difficult examples. The proposed Asymmetric Gentle AdaBoost takes account of the situation that the rare positive targets have to be distinguished from enormous negative patterns in practical applications. The quantitative evaluation on the well-defined INRIA data set demonstrates the effectiveness of our methods. © 2009 Springer Berlin Heidelberg.