Real-Time Attention for Robotic Vision

This thesis is concerned with the development and real-time implementation of an algorithm to determine interest points in a scene. These interest points will be used by a robot-mounted camera to focus its attention. The camera output has a nonuniform sampling resolution, modeled after the primate visual system. It provides a central high resolution foveal region surrounded by a much coarser peripheral region. The objective is to continuously position the camera so that the interesting areas in the scene lie within the foveal region. To this end, a computational model of visual attention has been developed and implemented in this thesis. The algorithm is based on psychophysical experiments of human gaze fixation. Using context-free edge information as input, interest points are defined as centres of regions surrounded by edges. This is shown to be equivalent to defining interest points as the points of intersection of lines of symmetry between edges in an image. By adopting a symmetry measure based on the loci of centres of cocircular edges, a novel, real-time method for computing these interest points is proposed. The algorithm has been implemented on a parallel network of Texas Instruments TMS320C40 (C40) processors. With this configuration, processing rates can exceed ten frames per second, depending on the algorithm parameters. The thesis also shows results of the algorithm applied to a wide range of real-world images, both foveal and peripheral, as well as an analysis of parameter sensitivity, and system throughput and latency.