A Real-Time Foveated Sensor with Overlapping Receptive Fields

Abstract

Vision systems for autonomous robots sometimes require high resolution, sometimes a wide field-of-view, and always fast processing. To achieve these same goals, the primate retina performs non-linear ‘image’ data reduction which ultimately produces a relatively small output. Such a data reduction scheme provides a compromise between the requirements of field of view, resolution, and speed of processing. An overlapping receptive field (RF) data reduction model, as proposed in the system presented here, is based on the retina, and offers flexibility in the selection of RF averaging masks. This flexibility is illustrated with a set of outputs produced using uniform, Gaussian, difference-of-Gaussians, and edge detection masks. However, such models are more computationally expensive than their non-overlapping counterparts. To compute the requisite mapping, an adapted scan-line algorithm is used due to its efficiency with respect to memory and speed. To achieve at least 10 frames/s, we employed MIMD parallel processing using six TMS320C40 digital signal processors. An image was captured by one processor, and the data was distributed line by line to up to four other mapping nodes. Each of the mapping nodes produced partial results that were combined by the last node. Throughput measurements showed that the minimum median throughput was 11.9 frames/s for useful model parameter combinations. Measurements using one to four mappings nodes showed that the speedup was linear with respect to the number of mapping nodes. The output of this data reduction system has been employed to compute points of interest in the field-of-view, which were then used to alter the camera gaze.