Estimating image velocity with convected activation profiles: analysis and improvements for special cases

The method of convected activation profiles was developed to measure short-range visual motion of edge and point features in time-varying imagery. Each feature is assumed to generate a spatiotemporal Gaussian activation profile that results in a shape-preserved activity wave that is convected along with that feature, and the phase velocity of the wave provides a velocity estimate of the feature. By this method, both explicit feature tracking (a complex and computationally expensive operation) and the assumption that intensity is convected (which is rarely justified) are avoided. The method is suitable for real-time implementations and can be described in terms of shunting dynamics of neural systems. Spatiotemporal filters that measure the velocity of lines and points were described and demonstrated in the earlier work: this paper presents a detailed analysis of the accuracy of the method in scenes consisting of highly textured objects with fixed projections onto the image plane. We also describe how to accurately measure the velocity of short lines and line ends; in the past the velocity of short lines was severely underestimated, and the velocity of line ends could only be measured by recognizing line end features and evaluating the speed of these “point” features in isolation. This new method simplifies velocity extraction yet requires no additional computation. Finally, we clarify our earlier suggestion for selecting a velocity estimate from among several filters of different scales