Specification, Estimation and Validation of a Pedestrian Walking Behaviour Model

Specification, Estimation and Validation of a Pedestrian Walking Behaviour Model


T Robin, M Bierlaire, J Cruz, Transp-or laboratory, EPFL, CH; G Antonini, Zurich IBM Research Center, CH


Understanding and predicting the pedestrians walking behaviour becomes important, especially in urban contexts. We propose and validate a model for pedestrian short range walking behaviour in normal conditions, based on a discrete choice approach.


Understanding and predicting the walking behaviour of pedestrians becomes more and more important, especially in urban contexts, where the density of people is higher and higher. In this paper, we propose and validate a model for pedestrian walking behaviour, based on a discrete choice approach. We are interested in modelling the short range behaviour in normal conditions, as a reaction to the surrounding environment and to the presence of other individuals. With the term normal we refer to non-evacuation and non-panic situations. The model has been estimated on a real dataset, and thoroughly validated using both the estimation dataset and another set of data not involved in the estimation process.

The model is specified as a discrete choice model, where the choice of the position of the next step is considered, based on a individual-specific discretisation of space, varying with walking speed and direction. The utility associated with each cell captures various behavioural patterns: willingness to reach the destination, stability of the direction, acceleration, interactions with other pedestrians. The latter pattern involves collision avoidance and leader-follower behaviours. The collision avoidance pattern is designed to capture the effects of possible collisions on the current trajectory of the decision maker. The leader-follower pattern is designed to capture the tendency of people to follow another individual in a crowd, in order to benefit from the space she is creating. The spatial correlation among the alternatives is taken into account defining a cross nested logit model.

The model has been estimated by maximum likelihood on a Japanese dataset, using Biogeme. The dataset consists of pedestrian trajectories manually tracked from video sequences. The subset has been collected in Sendai, Japan, on August 2000. Two main pedestrian flows cross the street, giving rise to a large number of interactions. In this context, 190 pedestrian trajectories have been manually tracked, for a total number of 9281 position observations. The estimated coefficients are significant and their signs are consistent with our behavioural assumptions. This model has been validated in order to evaluate its prediction capacity.

Our validation procedure consists in applying two models on two datasets. In addition to the model presented above, we consider also a simple model, where the utility of each alternative is represented only by an alternative specific constant (ASC). This model perfectly reproduces the observed shares in the sample. The objective is to illustrate the importance of the explanatory variables in the model. The two datasets are the Japanese one, and a dataset collected in the Netherlands. This second has been collected at Delft University, in the period 2000-2001. Volunteer pedestrians are called to perform specific walking tasks in a controlled experimental set-up, in order to create characteristic pedestrian motion patterns such as one-directional flow, bi-directional flow, walking through narrow and wide bottlenecks and crossing flows among the others. For the purpose of our validation procedure we used the subset of the Dutch dataset corresponding to a bi-directional flow. This situation is the experimental version of the Japanese dataset, which corresponds to a walkway. It includes 724 subjects for 47481 observed positions, collected by means of pedestrian tracking techniques on video sequences. The predictions on the Japanese dataset of the estimated model have been compared to those of the ASC model developed on this same dataset. In addition a cross-validation was performed on this dataset. As for the Japanese dataset the predictions of the estimated model have been compared to those of the ASC model on the Dutch dataset. This third part is the most powerful, because the Dutch dataset had not been involved in the model calibration. The proposed validation procedure underline a good stability of the model and a good generalisation performance. Few observations are badly predicted, mostly concentrated at the extreme of the choice set.

Finally, the model has been embedded in a simulator, allowing also for visual validation.


Association for European Transport