User trajectories contain a wealth of implicit information. The places that people visit, provide us with information about their preferences and needs. Furthermore, it provides us with information about the popularity of places, for example at which time of the year or day these places are frequently visited. The potential for behavioral analysis of trajectories is widely discussed in literature, but all of these methods need a pre-processing step: the geometric trajectory data needs to be transformed into a semantic collection or sequence of visited points-of-interest that is more suitable for data mining. Especially indoor activities in urban areas are challenging to detect from raw trajectory data. In this paper, we propose a new algorithm for the automated detection of visited points-of-interest. This algorithm extracts the actual visited points-of-interest well, both in terms of precision and recall, even for the challenging urban indoor activity detection. We demonstrate the strength of the algorithm by comparing it to three existing and widely used algorithms, using annotated trajectory data, collected through an experiment with students in the city of Hengelo, The Netherlands. Our algorithm, which combines multiple trajectory pre-processing techniques from existing work with several novel ones, shows significant improvements.