In this paper, a new methodology based on the Author Topic Model (ATM) method is presented to perform team activity recognition and analysis in handball videos. Instead of using players' trajectories we just rely on low level features related to local motion, the evolution of which is then modeled over time by the ATM. The proposed methodology is applied to the task of recognizing four kinds of team activities in handball videos from the CVBASE 06 dataset and to analyze which are the most important elements of the activities. Our method is compared with two other ways of characterizing videos based on Bag-of-Words (BoW) and Latent Dirichlet Allocation (LDA) techniques. Our proposal obtains competitive results in terms of accuracy, computing time and interpretation of the results.
© Copyright 2015 Neurocomputing. Elsevier. All rights reserved.