Analyzing activities in videos using latent Dirichlet allocation and granger causality

Abstract

We propose an unsupervised method for analyzing motion activities from videos. Our method combines Latent Dirichlet Allocation with Granger Causality to discover the main motions composing the activity as well as to detect how these motions relate to one another in time and space. We tested our method on synthetic and real-world datasets. Our method compares favorably with state-of-the-art methods.