New geometric descriptor for the recognition of aggressive human behavior

Video surveillance has become a ubiquitous feature of modern life that has the potential to detect and monitor aggressive behavior more accurately. In this paper, we propose a new paradigm to recognize an aggressive human behavior such as boxing action. This method has been carried out on two levels; the low-level analysis consists to characterize each frame by using a new geometric descriptor. The latter is associated with a bag-of-features approach in order to extract the local movement of actions. Each frame of atomic action is associated with an appropriate label. This is done by using an offline clustering algorithm such as k-means. The high-level analysis consists to generate the feature vectors from a sequence video by using a set of labels as an optimum codebook. The boxing actions are then recognized by applying a support vector machine classifier. The tests are conducted on our own database and KTH dataset actions. The obtained results show that the proposed method enables robust recognition of aggressive human behavior in very challenging situations such as dynamic environment and deals well with self-occlusion problem.