A velocity field approach to the detection of pedestrian interactions

The problem of detecting pedestrian interactions in image sequences is addressed. An extension of the motion field model to account for pedestrian interactions is presented. Pedestrian movements are modelled by a non-linear state model in discrete time, in which the trajectories follow a known field, in the absence of interactions, and include an extra term otherwise. The detection of interactions between pedestrians is reduced to a time varying parameter estimation problem, for which algorithms based on moving horizon estimation with a variable forgetting factor and on a Kalman filter are compared.