Sequential Dynamic Leadership Inference Using Bayesian Monte Carlo Methods

Abstract

Hierarchy and leadership interactions commonly occur in animal groups, crowds of people, and in vehicle motions. Such interactions are often affected by one or more individuals who possess key domain information (e.g., final destination, environmental constraints, and best routes) or pertinent traits (e.g., better navigation, sensing, and decision making capabilities) compared with the rest of the group. This article presents a framework for the automatic identification of group structure and leadership from noisy sensory observations of tracked groups. Accordingly, a new leaderfollower model is developed, which assumes the dynamics of the group to be a multivariate OrnsteinUhlenbeck process with the designated leader(s) drifting to the destination and followers reverting to the leaders state. Sequential Monte Carlo approaches, and specifically the sequential Markov chain Monte Carlo approach, are adopted to infer, probabilistically, the evolving leadership structure. A RaoBlackwellisation scheme is employed such that the kinematic state of the objects in the group is inferred in closed form by Kalman filtering. Experiments show that the proposed techniques can successfully determine the leadership structures in challenging scenarios with a corresponding enhancement in tracking accuracy through direct consideration of the leadership interactions of the group.