Efficient Optimization Algorithm for Space-Variant Mixture of Vector Fields

This paper presents a methodology for trajectory classification of human activities. The presented framework uses a mixture of non parametric space-variant vector fields to describe pedestrian’s trajectories. An advantage of the proposed method is that the vector fields are not constant and depend on the pedestrian’s localization. This means that the switching motion among vector fields may occur at any image location and should be accurately estimated. In this paper, the model is equipped with a novel methodology to estimate the switching probabilities among motion regimes. More specifically, we propose an iterative optimization of switching probabilities based on the natural gradient vector, with respect to the Fisher information metric. This approach follows an information geometric framework and contrasts with more traditional approaches of constrained optimization in which Euclidean gradient based methods are used combined with probability simplex constraints. We testify the performance superiority of the proposed approach in the classification of pedestrian’s trajectories in synthetic and real data sets concerning far-field surveillance scenarios.