A statistical surveillance system gives a signal as soon as data give enough evidence of an important event. We consider on-line surveillance systems for detecting changes in influenza incidence. One important feature of the influenza cycle is the start of the influenza season, and another one is the change to a decline (the peak). In this report we discuss statistical methods for on-line peak detection. One motive for doing this is the need for health resource planning. Surveillance systems were adapted for Swedish data on laboratory verified diagnoses of influenza. In Sweden, the parameters of the influenza cycles vary too much from year to year for parametric methods to be useful. We suggest a non-parametric method based on the monotonicity properties of the increase and decline around a peak. A Monte Carlo study indicated that this method has useful stochastic properties. The method was applied to Swedish data on laboratory verified diagnoses of influenza for seven periods.