Abstract

In this talk I will describe recent methods and applications for high-level inference and tracking of multiple object, groups and networks, by incorporation of behavioural interactions and unobserved intent into the dynamical models. The idea here is to extend the standard multiple object tracking paradigm to one in which we may automatically learn dynamic interactions between those objects, as well as infer possible intentionalities of the objects. Our models are based on principles from animal behavioural analysis in which objects follow patterns of behaviour based loosely upon what their neighbours in the group are doing, and upon the (unknown) intentionality of the group, for example its final destination. We may also learn more complex interactions such as whether one member of the group is a `leader’ of the dynamics and how the objects are split between different groupings. Models are typically formulated in continuous time, and inference is carried out on-line using numerical Bayesian filtering strategies, implemented with state of the art methods such as particle filters and Markov chain Monte Carlo. Applications will be presented from the areas of vehicle tracking, wild animal pack hunting behaviour analysis, financial time series, and finally applications in User Interfaces for automobiles in which the task is to determine accurately and rapidly the intended icon a user is pointing at on a screen, based on the trajectory of hand motion near to the screen, and in the presence of disturbances from suspension and road surface.