Can walking behaviour be predicted? An analysis of the calibration and fit of pedestrian models

In the last decades a number of different pedestrian movement simulation models have been proposed. The primary use of these models lies in the planning and evaluation of large pedestrian infrastructures such as transportation hubs with a focus to increase comfort and safety for pedestrians. While the number of different simulation models proposed is increasing at a fast pace not much is known on the properties of calibration procedures as well as the transferability of the models estimated in one setting to other settings. This paper compares three different calibration methods for a slightly adapted social force model. The main emphasis lies on the characteristics of the data generation process and the information contained in the data sets. Furthermore an investigation into the sensitivity of the model-parameters of the calibrated model is carried out and some tests on the transferability of the model to different scenarios are done. The main results in this respect are that data quality has a strong effect on the suitability of different calibration strategies and that the information content in the scene under investigation limits the transferability of the results to other scenarios. This leads to the suggestion that there is a need to include several data-sets with different characteristics in the calibration process to achieve a model that performs well on a wider variety of settings.