Wrong turn - No dead end: A stochastic pedestrian motion model

This paper addresses the use of social behavior models for the prediction of a pedestrian's future motion. Recently, such models have been shown to outperform simple constant velocity models in cases where data association becomes ambiguous, e.g. in case of occlusion, bad image quality, or low frame rates. However, to account for the multiple alternatives a pedestrian can choose from, one has to go beyond the currently available deterministic models. To this end, we propose a stochastic extension of a recently proposed simulation-based motion model. This new instantiation can cater for the possible behaviors in an entire scene in a multi-hypothesis approach, using a principled modeling of uncertainties. In a set of experiments for prediction and template-based tracking, we compare it to a deterministic instantiation and investigate the general value of using an advanced motion prior in tracking.