A time delay neural network algorithm for estimating image-pattern shape and motion

摘要

In this paper we present a novel concept for simultaneous shape estimation and motion analysis based on a feed-forward TDNN architecture with adaptable spatio-temporal receptive fields. On synthetic image sequences displaying elliptic spots of different orientation moving horizontally across the scene at several speeds, this network simultaneously manages to classify the shapes correctly as well as to estimate their speed and motion direction, given various test sets and network parameter settings. A very interesting feature is the property that a network having learned a certain number of shape and motion classes is able to generalize to intermediate shapes and speeds it has never `seen' during training by interpolating between the learned pattern classes. Moreover, the network turns out to be rather robust with respect to random deviations of the actual motion from the trained motion patterns. We furthermore apply the network successfully to a simple example of real-world data.