In this paper we learn patterns of activity in open urban spaces and detect activity outliers that represent events of interest. We do so utilising background suppression to flag people as foreground blobs in videos from city surveillance cameras. Since the application domain is challenging, with far-field cameras viewing scenes that vary from completely empty to very crowded, and each person in the crowds being a handful of pixels, we first establish the performance of different background subtraction algorithms using manually annotated scenes. We then apply the best-performing SubSENSE algorithm in off-line videos collected over many days, to learn the activity patterns and detect the events of interest as outliers.