An Expectation Maximisation Algorithm for Behaviour Analysis in Video
Abstract
Surveillance systems require advanced algorithms able to make decisions without a human operator or with minimal assistance from human operators. In this paper we propose a novel approach for dynamic topic modeling to detect abnormal behaviour in video sequences. The topic model de- scribes activities and behaviours in the scene assuming behaviour temporal dynamics. The new inference scheme based on an Expectation-Maximisation algorithm is implemented without an approximation at intermediate stages. The proposed approach for behaviour analysis is compared with a Gibbs sampling inference scheme. The experiments both on synthetic and real data show that the model, based on Expectation-Maximisation approach, outperforms the one, based on Gibbs sampling scheme.