Social networks attract lots of new users every day and absorb from them information about events and facts happening in the real world. The exploitation of this information can help identifying mobility patterns that occur in an urban environment as well as produce services to take advantage of social commonalities between people. In this paper we set out to address the problem of extracting urban patterns from fragments of multiple and sparse people life traces, as they emerge from the participation to social network. To investigate this challenging task, we analyzed 13 millions Twitter posts (3 GB) of data in New York. Then we test upon this data a probabilistic topic models approach to automatically extract urban patterns from location-based social network data. We find that the extracted patterns can identify hotspots in the city, and recognize a number of major crowd behaviors that recur over time and space in the urban scenario.