An evaluation of gaze modulated spatial visual search for robotic active vision

Type Conference Proceeding (Non-Journal item)
Original languageEnglish
Title of host publicationTAROS 2010
Subtitle of host publicationProceedings of Towards Autonomous Robotic Systems 2010
PublisherPlymouth University Press
Pages83 - 90
Number of pages8
ISBN (Print)978-1-84102-263-5
Publication statusPublished - Aug 2010
EventTowards Autonomous Robotic Systems 2010 - Plymouth, United Kingdom of Great Britain and Northern Ireland
Duration: 31 Aug 201002 Sep 2010

Conference

ConferenceTowards Autonomous Robotic Systems 2010
CountryUnited Kingdom of Great Britain and Northern Ireland
CityPlymouth
Period31 Aug 201002 Sep 2010
Links
Permanent link
View graph of relations
Citation formats

Abstract

Active vision is an essential part of many autonomous robot systems, in particular humanoid robots. In this work we present a method for spatial visual search which is modulated by the absolute motor positions of the active vision system resulting from saccades to objects. A central element of this approach is the so called visual memory where these motor configurations are stored. Based on these motor data, the system can evaluate which of the current visual stimuli have already been saccaded to. In this sense, motor configurations in the visual memory modulate the selection of visual targets for the purpose of saccade. Two architectures are presented which instantiate this gaze modulated visual search in a robotic scenario. The paper also presents a series of systematic experiments demonstrating the impact of two essential parameters (E [size of inhibitory neighborhood] and G [decay rate]) on the behavioural dynamics of the active vision system. E was found to determine the number of saccades needed to scan a scenario whilst G controlled the persistence of visual memory. Finally, we discuss the advantage of gaze modulated visual search compared to other common strategies without gaze-modulation. It is apparent that gaze space modulation is advantageous with respect to real-time performance and scalability, and therefore offers an interesting alternative approach for active vision in robotics as well as for general models of visual search.