Crowd analysis with target tracking, K-means clustering and hidden Markov models

The paper presents a framework for crowd analysis that can handle both sparse and dense crowds, by combining micro- and macroscopic crowd analysis approaches. The paper focuses on detection, tracking and behaviour of dense crowds. We use multiple target tracking (MTT), group tracking, K-means clustering and hidden Markov models (HMM). K-means clustering is used to decide if micro- or macroscopic approaches should be used. A first evaluation, based on recorded and simulated data sets, has been done. The evaluation shows that MTT works well when the crowd is relatively sparse. When the crowd becomes dense track identities are easily switched between tracks. For dense crowds centroid-based group tracking is proposed. The algorithms for dense crowd detection and behavior recognition show promising results. The accuracies of the algorithms range from 84 % and above. Increased internal crowd activities will, however, temporarily reduce the accuracy of the centroid-based group tracking.