Trajectory-based human activity recognition with hierarchical dirichlet process hidden Markov models

Trajectory-based human activity recognition aims at understanding human behaviors in video sequences. Some existing approaches to this problem, e.g., hidden Markov models (HMM), have a severe limitation, namely the number of motions has to be preset. In fact, this number is difficult to define in advance in real practice. To overcome this shortcoming, we propose a new method for modeling human trajectories based on the hierarchical Dirichlet process hidden Markov models (HDP-HMM), and adopt a Gibbs sampling algorithm for model training. Using our proposed technique, the number of motions can be inferred automatically from data and is also allowed to vary among different classes of activities. Experiments on both synthetic and real data sets demonstrate the effectiveness of our approach.