Identifying the Computational Requirements of an Integrated Top-Down-Bottom-Up Model for Overt Visual Attention within an Active Vision System

Abstract

Computational visual attention systems have been constructed in order for robots and other devices to detect and locate regions of interest in their visual world. Such systems often attempt to take account of what is known of the human visual system and employ concepts, such as ‘active vision’, to gain various perceived advantages. However, despite the potential for gaining insights from such experiments, the computational requirements for visual attention processing are often not clearly presented from a biological perspective. This was the primary objective of this study, attained through two specific phases of investigation: 1) conceptual modeling of a top-down-bottom-up framework through critical analysis of the psychophysical and neurophysiological literature, 2) implementation and validation of the model into robotic hardware (as a representative of an active vision system). Seven computational requirements were identified: 1) transformation of retinotopic to egocentric mappings, 2) spatial memory for the purposes of medium-term inhibition of return, 3) synchronization of ‘where’ and ‘what’ information from the two visual streams, 4) convergence of top-down and bottom-up information to a centralized point of information processing, 5) a threshold function to elicit saccade action, 6) a function to represent task relevance as a ratio of excitation and inhibition, and 7) derivation of excitation and inhibition values from object-associated feature classes. The model provides further insight into the nature of data representation and transfer between brain regions associated with the vertebrate ‘active’ visual attention system. In particular, the model lends strong support to the functional role of the lateral intraparietal region of the brain as a primary area of information consolidation that directs putative action through the use of a ‘priority map’.