Theories of optimal coding propose to understand early sensory processing of stimuli as being adapted to the statistics of the signals naturally occurring when interacting with the environment. Given that perception is active, the stimuli at the retina are dependent on oculomotor control and therefore on the executed task. How do different tasks affect the statistics of image features at the fixation location? In the past a number of studies have shown that the features at fixation locations are different from those selected randomly in natural scenes. Those results were obtained by subjects looking at briefly presented static images of natural scenes subtending a limited visual angle in a laboratory environment while mostly executing the so-called 'free-view' task. How do these results transfer from the laboratory environment to real natural environments in which subjects execute natural goal-directed behavior?

In order to address this question, eye movements of human subjects acting in a natural wooded environment were tracked with a custom-made portable eye tracker. An HD-video camera mounted on a bicycle helmet recorded the visual scene with a diagonal field of view of 75 degrees.

This high-resolution video allowed for the accurate analysis of visual features at fixation location versus locations chosen randomly within the visual scene. Three subjects each executed a search task, a 'free-view' task and a walking task. Each task had duration of approximately 90 seconds. The eye tracker was calibrated between trials.

Here we report results from the analysis of the spatiotemporal autocorrelation functions obtained from the natural scene sequences separately for the three different tasks and sets of image patches chosen according to three random sampling methods. The data show significant statistical differences between the tasks. We therefore conclude that theories of optimal coding of sensory stimuli should take into account explicit task dependencies.