Abnormal behaviour detection in video using topic modeling

The growth of the number of surveillance systems makes it is impossible to process data by human operators thereby autonomous algorithms are required in a decision-making procedure. A novel dynamic topic modeling approach for abnormal behaviour detection in video is proposed. Activities and behaviours in the scene are described by the topic model where temporal dynamics for behaviours is assumed. Here we implement Expectation-Maximisation algorithm for inference in the model and show in the experiments that it outperforms the Gibbs sampling inference scheme that is originally proposed in [1].