Fig. 13a shows a histogram of group-specific travel times according to pedestrian tracking data and as computed by Ped-
CTM. As defined previously, a pedestrian group comprises all pedestrians that embark on the same route within the same
departure time interval. The bin size corresponds to the simulation time step, i.e., to Dt ¼ L=vf ¼ 2:22 s. PedCTM slightly
overestimates the frequency of travel times beyond 70 s, and underestimates the occurrence of short travel times below 20 s.
To allow for a comparison with the social force model, Fig. 13b provides a similar comparison of route-specific travel
times with a temporal aggregation of 60 s. The resulting mean values are shown in a histogram with a bin size of 5 s. As seen
previously, PedCTM overestimates the occurrence of long travel times, but otherwise shows a reasonable agreement with
trajectory data. The social force model seems to overestimate small travel times and to underestimate the most frequently
observed travel times in the range between 35 and 55 s. Overall, the agreement between PedCTM and trajectory data is better
than for SFM if the squared error is considered (114.4 vs. 181.6).
Fig. 14 shows the residuals of group-specific travel times computed using PedCTM compared to tracking data. The distribution
of the absolute and the relative error is provided in Fig. 14a and b, respectively. In a first approximation, the histograms
resemble a normal distribution with zero mean. From Fig. 14a, it can be seen that the predicted travel times deviate
less than 10 s from the observed walking times for two thirds of the pedestrian groups. For less than 10% of all groups, the
deviation is larger than 20 s. In relative terms, 50% of the estimates deviate less than 13% from observed values, and for more
than 80% of all groups the relative error is smaller than 33%. Also, it can be seen from the absolute residuals that there are a
few particularly long walking times which cannot be reproduced by PedCTM at all.
Another way of illustrating the performance of PedCTM in predicting travel times is by means of a scatter plot.
Fig. 15a shows observed versus predicted travel times in PU West, again aggregated by routes and 60 s-intervals. At first
sight, the predictive power may seem low. However, a large majority of the travel time-pairs come to lie reasonably
close to the 45 best-fit line (dashed curve in Fig. 15a). Specifically, for 74% of all observations, the relative error
amounts to less than 25%. For around 15% of all observations, PedCTM underestimates the walking time by more than
25% (see observations in the lower right half of Fig. 15a, or the left tail of the distribution in Fig. 14a). A reason for this
discrepancy might be non-walking behavior of pedestrians, such as purchasing a ticket or checking the train timetable.
These activities are not considered by the pedestrian walking model. Similarly, PedCTM fails at predicting very short
travel times corresponding to average velocities that are higher than the free-flow speed (right tail in Fig. 14a and
b). Both shortcomings are due to the assumption of a deterministic fundamental diagram, which associates with each
density exactly one speed. We believe that these limitations could be overcome by incorporating a stochastic fundamental
diagram (Nikolic´ et al., 2014) as well as by means of a multi-class framework (Cooper, 2014).
Showing a mixture of routes with different lengths, it could be argued that Fig. 15a should not be used to assess the ability
of the model to predict traffic-dependent variations in travel times. If only a single route is considered, then this dependency
is explored more readily. This may be done at the example of the Dutch bottleneck experiment, which shows a particularly
high level of demand–supply interaction. Fig. 15b shows the corresponding travel time scatter plot, where a temporal aggregation
of Dt ¼ L=vf ¼ 0:57 s is used (aggregation by pedestrian groups). Data points used for calibration and validation are
shown separately. A variation in travel time by a factor of 5, from 5 s to 25 s, is discernible. Since there is only a single route,
this variation is solely a consequence of changing traffic conditions. For a large majority of data points, PedCTM is capable of