The detection of security-relevant events in large crowds is a difficult vision problem. We investigated differences in visual search of dangerous events between security experts and naive observers during the observation of large scenes, typically encountered on the grandstand of stadiums during soccer matches. Based on a new algorithm for the synthesis of crowd scenes with well-controlled statistical properties, subjects were eye-tracked during the observation of such scenes. Detection rates, fixation rates and times were assessed from 12 naive subjects and a few expert observers. The results were compared with two theoretical models: (1) an approach for saliency computation based on low-level features (Bruce and Tsotsos, 2009), and (2) a graphical Bayesian model that was trained with expert knowledge derived from scenes containing security-relevant events, exploiting optic flow features extracted with a neurally plausible algorithm (Beck et al, 2007). We characterize the relative importance of salience and expert knowledge for the generation of correct detections and the visual search strategies for both types of observers. We found substantial differences betwen the observer types, suggesting a specific knowledge-driven search strategy for the expert observers.