The purpose of this chapter is both to review some of the most representative visual attention models, both theoretical and practical, that have been proposed to date, and to introduce the authors’ attention model, which has been successfully used as part of the control system of a robotic platform. The chapter has three sections: in the first section, an introduction to visual attention is given. In the second section, relevant state of art in visual attention is reviewed. This review is organised in three areas: psychological based models, connectionist models, and features-based models. In the last section, the authors’ attention model is presented.
TopVisual Attention
Attention can be defined as the cognitive process by which human beings focus on a certain aspect of the environment while ignoring all others, or as the management of the resources that allow for information processing (Anderson, 2004). Talking to someone while ignoring other conversations around or driving properly while at the same time ignoring distracting elements adjacent to the road are examples of attention. Attention is one of the most studied topics in psychology and neuroscience.
Attention may be influenced by many channels of information (like animals obtain information from their different senses), referring to the case when attention works with visual information channels as “visual attention”. Since this is the only source of information considered in this chapter, even when we refer to it generically as “attention” it will always be visual in nature.
There are several reasons why attention is useful for any system, whether biological or artificial. The most obvious is to reduce input data thus freeing up resources used, resources which could then be employed in other system tasks. It is also interesting the ability to simplify a complex problem into simpler ones. Additionally, using a mechanism of attention, the influence of information that may cause distractions can be suppressed or at least reduced.
Attention in Robotics
Robotic vision systems are often structured as a system of analysis at one end and a response system in the other. The analysis part generates a model of the environment using the images obtained from the same and the response part uses these models to plan or perform an action. Vision in this case acts as a pre-action stage, calculating all the features that the planning system might need.
Thanks to attention, instead of processing all pixels in each image, with the amount of resources that would imply, only those which presumably contain information relevant to the current task are selected. That is, the system adapts to extract from the images only the information it really needs.
Attention seen as a means to select information for tasks, as if it were a filter, is also very useful in the field of robotics. Autonomous robots should be able to interact with complex environments and scenarios, simultaneously maintaining different goals for each of their possible behaviours, the latter being understood as sets of actions that seek a goal.
Attention in this case can guide robots behaviour, restricting the amount of processed visual information to what is relevant to the tasks at hand and ignoring other elements (distractors). It also provides a coordination mechanism, because selection of stimuli can serialise the actions of the current behaviour.
The latest trends in robotics to solve problems related to attention go through the application of bio-inspired models (Frintrop et al., 2005a; Itti & Koch, 2000; Navalpakkan & Itti, 2006; Torralba et al., 2006; Tsotsos et al., 1995). It is important to note that although algorithms used are biologically inspired their goal is not to model biological systems.