Trajectory Recognition Based On Variational BP-HMMs

Trajectory recognition is to build a classification model based on labelled trajectories and then use the model to predict the classes of unlabelled trajectories. The hidden Markov model (HMM) is a vital model for trajectory recognition. As the hidden state number of the HMM is important and hard to be determined, many methods like Beta process hidden Markov models (BP-HMMs), hierarchical Dirichlet process hidden Markov models (HDP-HMMs) and hidden Markov models with Bayesian information criterion (HMM-BIC) have been proposed to solve it. However, among these methods, only the BP-HMM can find the sharing patterns among trajectories, which has been proved helpful to human activity recognition and useful in other trajectory recognition scenes. The existing approximate inference for the BP-HMM is based on the sampled method which has some inherent defects, e.g., the difficulty of justifying the convergence time. In order to get an effective method to learn the number of hidden states and the shared scheme automatically, we propose variational Bayesian BP-HMMs for trajectory recognition. In the variational BP-HMM, the Beta process (BP) is constructed by stick-breaking process which is amenable to variational inference. Further, in order to get a more efficient and natural method, we reconstruct an HMM for each class of trajectories instead of an HMM for every trajectory. In this paper, we show that the variational BP-HMM which builds an HMM for each class is more efficient and has better performance. Our method has great performance in the applications like the navigation task with mobile robot, handwritten character recognition and human activity recognition.