Pairwise Kernels for Human Interaction Recognition.

In this paper we model binary people interactions by forming temporal interaction trajectories, under the form of a time series, coupling together the body motion of each individual as well as their proximity relationships. Such trajectories are modeled with a non-linear dynamical system (NLDS). We develop a framework that entails the use of so-called pairwise kernels, able to compare interaction trajectories in the space of NLDS. To do so we address the problem of modeling the Riemannian structure of the trajectory space, and we also prove that kernels have to satisfy certain symmetry properties, which are peculiar of this interaction modeling framework. Experiment results show that this approach is quite promising, as it is able to match and improve state-of-the-art classification and retrieval accuracies on two human interaction datasets.