Control of the locus of gaze and deployment of attention in natural environments is largely unexplored. Although previous studies have revealed the task-directed nature of fixations, it is not known how this interacts with the stimulus properties of the image. Any task-driven, or top-down system must deal with the issue of when the particular computations are carried out. The observer must somehow access the perceptual information that is not on the current agenda. How does this selection occur? We tested the hypothesis that attentional control is determined primarily by top-down factors, and that the distribution of attention is sensitive to the expectations the observer has about the current environmental context. Subjects walked along a footpath in a virtual environment. Fixation patterns were examined to measure sensitivity to virtual pedestrians on a collision path. Occasionally (~10% probability), a pedestrian on a non-colliding path changed onto a collision course for about 1 sec, after which it returned to its original, non-colliding path. If observers are continuously monitoring peripheral vision, or if the deviation attracts attention, this should be revealed by a fixation, or by avoidance, or by both. Subjects fixate non-colliding pedestrians about 73% of the time, with a latency of 740msec after they appear in the visual field. These pedestrians are fixated a second time with about 30% probability. When a pedestrian deviates onto a collision path, subjects fail to fixate the collider on 41% of trials. When subjects do make a fixation onto the collider, the latency between the path deviation and fixation was 548msec on average. The high probability of failing to fixate a collider suggests that it does not attract attention unless the observer is actively monitoring for pedestrians in peripheral vision. This is consistent with models where attentional control is determined by the task, with a minimal role for bottom up factors (Ballard et al, VSS 2003).