Skip to main content Skip to main navigation


Trace Driven Simulation Model for City Scale Crowd Movements

Tobias Franke; Andreas Poxrucker; Gernot Bahle; Paul Lukowicz
In: Proceedings of the 14th International Conference on Smart City. IEEE International Conference on Smart City, 14th, December 12-14, Sydney, Australia, IEEE, 2017.


We present a city-scale crowd simulation model based on a large data set (25 million GPS data points from 28’000 volunteers recorded during a 3-day city-wide festival held in Zurich in 2013). The model is based on a spatio-temporal abstraction of the festival, focusing on event sites and event times. Thus, we assume a certain number of events (concerts, shows, etc. as it’s typical at such festivals) taking place at different locations to be the key factor in how the majority of people move around the city. We then divide the city into cells characterized by their relation to such events (e.g. street leading to an event, event location etc.). For each cell, we separately consider different temporal phases of each event (people arriving, event itself, people leaving, global city-wide events such as fireworks, etc.) and, for each phase, derive the crowd motion characteristics from our data set. In the simulation, events can be placed on the city map at different times with cell types being defined accordingly. We then simulate the overall city-wide crowd motion/distribution by assigning the respective crowd motion characteristics to each cell at the respective event times. This allows the insights gained from one concrete festival data recording to be transferred to various festival settings and city layouts. This can be used for planning future events as well as for simulative evaluation of various smart city services (e.g. effectiveness of various P-2-P information delivery or collection strategies). We evaluate our model by considering different festival days (which have different event layouts) from our data set and show that simulations generated on the basis of our abstraction have good correspondence to the actual observed crowd distribution.