Pedestrian detection based on background modeling and head-shoulder recognition

Pedestrian detection is of much importance for its practical applications. This paper develops a novel pedestrian detection system which consists of three stages: motion region detection based on background modeling, feature extraction in the guidance of prior information, and map-based classification applying support vector machine (SVM) and Adaboost. First of all, an adaptive Gaussian Mixture Model is proposed to reduce the search for human targets in the background region. Secondly, the paper extracts a variant of HOG (Histograms of Oriented Gradients) and Haar-like feature to describe pedestrians, according to the prior information of human's relatively stable head-shoulder structure in various views. Thirdly, for the best performance of feature descriptors, this paper applies the combination of SVM (Support Vector Machine) and Adaboost, separately for HOG and Haar-like feature, as the final classifier. Experiment results validate the effectiveness of our method.