A Foveated Retina for Robotic Vision

A Real-Time Foveated Sensor with Overlapping Receptive Fields

“…To constrain the placement of these centres, we need to derive two positioning parameters from the three model parameters, namely, the angle between adjacent rays and the ratio k of radii of adjacent circles. The ray spacing angle is given by (1) As for the radii ratio, we used an adjacent circle spacing constraint which ensures that the average of the RF overlap of an RF on a circle i by its neighbors on circles i Ϫ 1 and i + 1 is the desired overlap . Given that the first ring of RFs is at the foveal radius r f (where i = 0), this constraint forces the circles to have radii r i , given by The difference between these positioning parameters and Yamamoto's version is very small (see [1] for the derivations of k and and details of this comparison).…”

“…The high output rate of the camera makes this very difficult. Yet analysis shows that using the SL approach requires fewer computations to compute the mapping than a straightforward table look-up approach [1]. The SL algorithm is described here in the context of computing the retino-cortical mapping.…”

“…A theoretical comparison between the ASL algorithm and one based on a per pixel/per RF look-up table technique shows that the ASL algorithm requires 65% less memory and 58% fewer processing cycles [1].…”

“…The basis of this compromise is the use of a variable resolution sensory system as an image data reduction scheme (see [1] or [2] for details of the biological facts behind this scheme). This technique produces two output images, as illustrated in Figure 1.…”

Attentive Visual Motion Processing: Computations in the Log-Polar Plane

“…We will not delve in this issue here. It has been extensively studied in [3] but it still remains an open problem. In our implementation we used non-overlapping averaging receptive elds as implemented in the emulation of the space-variant sensor in [22].…”