New paradigm for recognition of aggressive human behavior based on bag-of-features and skeleton graph

The human action recognition is an active field on computer vision in the last decade. It consists to automatically identify of human behavior and interpreting their actions. In this paper, we propose a new paradigm to recognize aggressive human behavior based on two models. The first method is based on shape representation by using bag-of-features approach and the second method is based on the skeleton graph in order to extract motion features. The feature association of the two models is carried out at each frame of atomic action. Thus, an appropriate label is assigned to each feature association vector by using an offline clustering algorithm such as k-means. The obtained feature vectors are conducted from a sequence video by using a set of labels as an optimum codebook. The aggressive behaviors are then recognized by applying a support vector machine classifier. The proposed algorithm enables robust recognition in very challenging situations such as dynamic environment and deals well with self-occlusion problem. Experimental results are conducted on KTH dataset actions and demonstrate that the proposed approach provide significant recognition rate of 96%.