Learning Generative Socially Aware Models of Pedestrian Motion

Abstract

Among the biggest challenges faced when integrating robots and intelligent vehicles in crowded environments is the difficulty encountered by artificial agents in anticipating people's moves, which is indeed a critical machine reasoning task in order to enable safe and comfortable interaction with humans. In general terms, this is a multi-agent prediction problem, whose solution, from a Bayesian perspective, involves on one hand inferring the joint latent dynamical state of all agents conditioned on observed data and on the other hand learning the corresponding system dynamics. Among the advantages of such a probabilistic approach is the fact that it naturally allows to account for the intrinsic stochasticity of both human behavior and sensor data in a principled manner. In this letter, we propose a novel Bayesian generative model, which generalizes a popular class of deterministic multiagent motion models and is capable of encoding traffic agents' interactions in a latent probabilistic space. We demonstrate, using real pedestrian data-sets, how it can be exploited for inference, prediction, and trained for learning of agent interaction dynamics. Our experimental evaluations indicate that the proposed approach outperforms both probabilistic latent variable models that neglect multi-agent interactions as well as deterministic physics-inspired models of human motion in crowds.