Behavioral modeling at different scales for trajectories gathered from cluttered scenes. Non parametric Bayesian approach using HDP

This article addresses the problem of behavioral modeling based on trajectories extracted from crowded scenes, especially in the situation of limited knowledge on the data. We propose to use a nonparametric Bayesian approach - the MLC-HDP defined by D. Wulsin - to discover and model reoccurring behaviors at different time scales. Thus, the exploratory analysis of trajectories is performed by an unsupervised classification, simultaneously over different semantic levels, with the number of clusters for each level not defined a priori but estimated from the data. Firstly, we validated our approach using a pseudo-ground truth generated using a multi-agent system which is able to simulate, with relatively low a priori, distinct behaviors and associated trajectories. Secondly, we tested our approach on real honeybee trajectories. © 2015 Lavoisier.