Neurodynamical model for visual action recognition

Abstract

The recognition of body motion requires the temporal integration information over time. This integration likely is achieved by dynamic neural networks, which are composed from neurons that are selective for form and optic flow patterns [1]. Physiological evidence also indicates that many action-selective visual neurons are view-dependent, so that such representations likely represent not only the time structure of actions, but also the stimulus view. The dynamic and self-organization properties of such representations have rarely been studied. We propose a neurodynamical model for the visual encoding of actions that is based on a two-dimensional neural field with the defining equations τuu.(φ,θ,t)=-u(φ,θ,t)+w(φ,θ)*[u(φ,θ,t)]++s(φ,θ,t)-αa(φ,θ,t)+ξ(φ,θ,t)τaȧ(φ,θ,t)=-a(φ,θ,t)+[u(φ,θ,t)]+ u specifies the membrane potential and a the adaptation state. The recurrent interaction kernel w is symmetric with respect to the origin in o-direction and asymmetric in ξ-direction, with an additional strong inhibitory component. The spatial convolution * is periodic. The stimulus signal s models an (idealized) activity distribution of shape- (or optic flow-) selective neurons that are maximally responding to stimulus frame θ and view angle o of an ongoing action stimulus. The noise variable ξ is defined by a Gaussian process whose kernel was fitted in order to reproduce coarsely the correlation statistics, dependent on the tuning similarity of the neurons. Time scales and adaptation strength are specified by the positive constants τu, τvand α. The first equation defines a dynamic neural field that stabilizes a stimulus-locked travelling peak solution in ξ-direction, and a winner-takes-all competition in o-direction. The second equation specifies a simple adaptation process.