Dynamic Hierarchical Dirichlet Process for Abnormal Behaviour Detection\n in Video

This paper proposes a novel dynamic Hierarchical Dirichlet Process topic\nmodel that considers the dependence between successive observations.\nConventional posterior inference algorithms for this kind of models require\nprocessing of the whole data through several passes. It is computationally\nintractable for massive or sequential data. We design the batch and online\ninference algorithms, based on the Gibbs sampling, for the proposed model. It\nallows to process sequential data, incrementally updating the model by a new\nobservation. The model is applied to abnormal behaviour detection in video\nsequences. A new abnormality measure is proposed for decision making. The\nproposed method is compared with the method based on the non- dynamic\nHierarchical Dirichlet Process, for which we also derive the online Gibbs\nsampler and the abnormality measure. The results with synthetic and real data\nshow that the consideration of the dynamics in a topic model improves the\nclassification performance for abnormal behaviour detection.\n