A treatise on log-polar imaging using a custom computational sensor

Gregory Lee Kreider, University of Pennsylvania

Abstract

Computational sensors begin to process transduced information directly on the sensing plane by integrating analytical circuitry. Their early calculations reduce the computational burden on later processors. Image processing in particular is a data-intensive problem that benefits from this relief. The Retina Imager is a custom CCD camera that samples on a spatially variant grid, which decreases the flow of visual data. Its sampling performs a log-polar transformation, decoupling the angular and radial axes of the polar plane. Although the resolution of the sensor drops at greater radial distances, it still has a region that can detect fine detail. Because the Retina imperfectly gathers data, it must respond to its environment, directing its gaze to important features. Although this active approach takes time to interact with the world, the advantages of the mapping (the decoupling and reduction of data) net a gain. The Retina, with its unique layout of expanding circles, does collect log-polar images and does demonstrate these gains. Edge and line detection, tracking, and time to impact estimates approach real-time speeds on a general purpose workstation using this camera, substantially outperforming similar algorithms running on standard television pictures. Simple low-level functions for controlling an active camera, such as measuring focus, also execute quickly. A sensor not locked to a conventional standard opens new feedback paths for an active system; varying the frame rate, for example, changes the perceived brightness of a scene. The Retina does improve the overall performance of an imaging system. Such focal plane image processing must draw on a wide range of analytical structures. New CCD implementations for signed arithmetic and mask convolution are useful low-level calculations. In a CCD/CMOS technology new devices for A/D conversion, neighborhood reconstruction, and line detection offer additional computational savings. Continued improvements in sensing hardware and active imaging, separately and together, will bring closer a robust comprehensive vision system. Foveated sensors, including the Retina, will be a fundamental part of those systems.

Subject Area

Electrical engineering|Computer science

Recommended Citation

Kreider, Gregory Lee, "A treatise on log-polar imaging using a custom computational sensor" (1993). Dissertations available from ProQuest. AAI9331807.
https://repository.upenn.edu/dissertations/AAI9331807

Share

COinS