The installation of surveillance networks has been growing exponentially in the last decade. In practice, videos from large surveillance networks are almost never watched, and it is frequent to see surveillance video wall monitors showing empty scenes. There is thus a need to design methods to continuously select streams to be shown to human operators. This paper addresses this issue and make three main contributions: it introduces and investigates, for the first time in the literature, the live stream selection task; based on the theory of social attention, it formalizes a way of obtaining some ground truth for the task and hence a way of evaluating stream selection algorithms; and finally, it proposes a two-step approach to solve this task and compares different approaches for interestingness rating using our framework. Experiments conducted on 9 cameras from a metro station and 5 hours of data randomly selected over one week show that, while complex unsupervised activity modeling algorithms achieve good performance, simpler approaches based on amount of motion perform almost as well for this type of indoor setting.