Credal human activity recognition based-HMM by combining hierarchical and temporal reasoning

Human activities recognition in videos sequences is a very current research topic being investigated in computer vision. This paper offers an approach for video analysis by exploiting hidden Markov models. We propose an extension of the standard model by integrating three abstraction layers through the management of hierarchical structure and the temporal evolution of events. In addition, data imperfections are also managed through a more generic framework than the probabilistic that is the Transferable Belief Model. The proposed approach has been assessed with the "baggage abandoned" scenario of PETS´06 dataset of computer vision community. Lastly, the proposed scenario recognition system performance is analysed and compared to the result of classic HMM models.