Boosted String Representation and Its Application to Video Surveillance

This paper presents a new behavior classification system for analyzing human movements directly from video sequences. First of all, we propose a triangulation-based method to transform each action sequence into a set of symbols. Then, for analyzing the human behavior via those strings representation, we propose a boosted string representation method to extract important string features for accurately analyzing and recognizing different action sequences. The boosted method not only can solve the problem of time warping, but also can reduce the error effects when some postures are wrongly coded into symbols. Since the Adaboost algorithm is proposed for solving two-class problems, we use the error coding concept to modify the Adaboost algorithm such that multiple human action events can be well solved. Then, each action can be well recognized by its correspondence boosted classifier. Experiment results prove that the proposed method is a robust, accurate, and powerful tool for human movement analysis.