To address the motion patterns learning task in crowded scenes

we propose a novel Distance Dependent Chinese Restaurant Franchise (DDCRF) mixture model

which is a hierarchical nonparametric Bayesian model based on dependencies DDCRF can learn the latent patterns accurately by introducing the distance information which is dependent of the observations of data points. We detail the generative process and Gibbs sampling process of DDCRF

and then the results of experiments are shown. An extensive evaluation is performed on the dataset including 47

866 tracklets collected from the crowded New York Grand Central station

indicating that our algorithm has the following advantages: deducing the number of latent motion patterns automatically

learning motion patterns precisely from these tracklets

and sharing constituent parts and sub patterns among different motion patterns.