Discovering Emergent Behaviors from Tracks Using Hierarchical Non-parametric Bayesian Methods

In video-surveillance, non-parametric Bayesian approaches based on a Hierarchical Dirichlet Process (HDP) have recently shown their efficiency for modeling crowed scene activities. This paper follows this track by proposing a method for detecting and clustering emergent behaviors across different captures made of numerous unconstrained trajectories. Most HDP applications for crowed scenes (e.g. traffic, pedestrians) are based on flow motion features. In contrast, we propose to tackle the problem by using full individual trajectories. Furthermore, our proposed approach relies on a three-level clustering hierarchical Dirichlet process able with a minimum a priori to hierarchically retrieve behaviors at increasing semantical levels: activity atoms, activities and behaviors. We chose to validate our approach on ant trajectories simulated by a Multi-Agent System (MAS) using an ant colony foraging model. The experimentation results have shown the ability of our approach to discover different emergent behaviors at different scales, which could be associated to observable events such as "forging" or "deploying" for instance.