Moving horizon estimation of pedestrian interactions using multiple velocity fields

This paper describes a model, and algorithms based on it, for the analysis of pedestrian interactions in outdoor scenes. Pedestrian activities are described by their trajectories in the scene, and we wish to know if they were independently generated or if they are correlated. Two models are considered: (i) a model based on multiple velocity fields recently proposed in Nascimento et al. (IEEE Trans Image Process 22(5):1712–1725, 2013) and (ii) an interaction model based on the attraction/repulsion between pairs of pedestrians. Several combinations of these models are studied and evaluated. An estimation method based on a moving horizon optimization of a quadratic cost functional is proposed. Experimental results with synthetic data and real video data are presented to assess the performance of the algorithms.