Learning a Social Force Model for Pedestrian Motion Analysis from Image Sequences

摘要

We propose a method for recursively learning the parameters of a numerical simulation model for pedestrian motion using an image sequence. We construct the model with so-called social forces, which have been successfully used in computer simulations for pedestrian motion analysis. The contribution of this paper is to combine the numerical simulation model and observations captured from image sequences. To this end, we introduce the framework of data assimilation, which is originally developed in geosciences such as weather forecasting and hydrology for refining numerical simulation models using observations available in the real world. In addition we use a particle filter for the recursive Bayesian estimation In experiments with real yideos we show a case study of pedestrian motion analysis.