Authors: [tex2html_wrap4222]M. Bolduc, M. D. Levine.
Investigator username: levine
Subcategory: sensor and processor design
The design of a vision system for autonomous robotics produces conflicting requirements. High resolution in the image is needed to obtain the necessary details in the local region of interest. But a wide field-of-view is also useful to detect events in the larger environment. For a uniform sensor, these entail a very large amount of image data, which in turn requires a large amount of processing time. However, to ensure adequate response time from the robot vision system, processing time must be kept at a minimum.
A good compromise between the resolution, field-of-view, and processing time requirements is to develop a camera system based on models of image data reduction of the primate retina. These non-uniform image sampling models effectively divide the retina into foveal and peripheral regions. The fovea is located at the center of the retina and contains image data at the highest possible resolution. In the periphery, the image information is compressed at a scale which varies inversely with the distance from the center of the fovea. This arrangement permits sensing a high level of detail in the area about the point of retinal fixation. At the same time it also gives coarse information about events in a large area around the center of attention without the cost of processing an extraneous amount of data.
An electronic retina (CCD camera with uniform resolution) is used as the input device. The foveation process is achieved by employing an adaptation of a raster-scan algorithm implemented on a network of parallel processing modules based on the Texas Instruments TMS320c40 DSP. As output, the system provides foveal and the peripheral images which will be fed directly to another processing system for further computation. The robot eye will be mounted on a specially designed miniature computer-controlled pan/tilt mechanism.