Conventional surveillance sensors suffer from an unaviodable tradeoff between image resolution and field of view. This problem may be overcome by combining a fixed, preattentive, low-resolution wide-field camera with a shiftable, attentive, high-resolution narrow-field camera. Here we present techniques for orienting the attentive camera to faces detected in the pre-attentive wide-field image stream.' Unfortunately, the low image resolution of the wide-field sensor precludes the use of most conventional face-detection algorithms. Instead, we argue that reliable performance can best be achieved by accurate probabilistic combination of multiple cues: skin detection, motion detection and foreground extraction. Fast sampling of scalespace over all three modalities is achieved using integral images and parametric models of response distributions are derived using supervised learning techniques. Log likelihood ratios for each modality are combined with spatial priors incorporating tracking and novelty objectives to yield a posterior map indicating the probability of a face appearing at each image location. The result is a real-time attentive visual sensor which reliably fixates faces over a 130 deg field of view, allowing high-resolution capture of facial images over a large dynamic scene.