Foveal Machine Vision Systems
Abstract
This work presents a new class of active machine vision systems, called foveal machine vision systems, which feature space variant sampling directed by gaze strategies. Two families of space variant sampling geometries are analyzed with spatial resolution decreasing with distance from the optical axis. One family features a linear acuity roll-off, and the other an exponential roll-off. Techniques are presented for the integration of sensor frames into unified static scene perceptions. Foveal systems can use many existing hierarchical processing techniques, in particular image pyramid structures and algorithms. A hierarchical structure called the foveal polygon is described. The foveal polygon is the subset of an image pyramid supported by foveal sensor frame. Top-down (coarse-to-fine) algorithms processing polygon data serve as drivers for gaze control. Additional gaze control strategies are presented for general learning and surveying (minimization of hypothesis entropy), and feature interrogation (hypothesis likelihood maximization).
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 01, 1990
- Accession Number
- ADA226274
Entities
People
- Cesar Bandera