Human Activity Recognition in Video Using a Hierarchical Probabilistic Latent Model

Abstract

In this work, we address the recognition of human activities from a sequence of visual data. To this end, a novel hierarchical probabilistic latent (HPL) model is proposed, which consists of four layers from bottom-up: spatiotemporal visual features layer, atomic pattern layer, latent topic layer, and behavior pattern layer. In this manner, the complicated human activities can be decomposed into low level features, atomic patterns, and latent topics, which are much better suited for the automatic understanding of human behaviors. Given a video sequence, both spatial and temporal interest points are extracted as the low level visual features, which are clustered into distributions of atomic patterns using hierarchical Bayesian networks (HBNs). Then, the proposed hierarchical probabilistic latent model is applied to represent the behavior patterns and latent topics as distributions over atomic patterns. Extensive experimental results based on the KTH dataset have demonstrated the efficiency of the proposed framework.