A Computer Vision–Enriched Discrete Choice Model for Pedestrian Route Choice Preferences in the Netherlands

Abstract

Walking is an important part of urban mobility in the Netherlands, but pedestrian route choice is often simplified in transport models by assuming that pedestrians choose the shortest or fastest route. This may overlook street-level characteristics that influence how people experience a walking route, such as greenery, traffic, pedestrian space, and the presence of other people. These factors may also matter differently depending on why someone is walking. This thesis therefore investigates to what extent visual and non-visual route attributes influence pedestrian route choice preferences in the Netherlands, and how these preferences differ across trip-purpose contexts.
To study this, a stated preference experiment was designed in which respondents repeatedly chose between two walking route alternatives. The alternatives differed in travel time and in the visual appearance of the street, shown through a street-level image. The choice tasks were presented in four contexts: walking to public transport, walking to work or school, walking home, and walking in free time. The collected data were used to estimate three model specifications for each trip purpose: a baseline Multinomial Logit (MNL) model with only travel time, a pixel-share MNL model with predefined visual attributes extracted from the images, and a Computer Vision-Enriched Discrete Choice Model (CV-DCM) that learns visual information directly from the full image.
The results show that travel time remains important, especially for more goal-oriented trips. However, travel time alone does not fully explain the observed choices. Models that include visual information generally perform better than the baseline model, with the CV-DCM showing the strongest predictive performance in most contexts. In the pixel-share MNL, visible greenery has the clearest and most consistent positive effect on route preferences. The qualitative validation of the CV-DCM also suggests that images with higher predicted utility are often greener and more attractive. At the same time, the results show that the role of visual information differs across trip purposes.
The trip-purpose-specific CV-DCM models were also applied in an exploratory pedestrian network analysis in Amsterdam-Zuid. Image utilities were linked to street segments and combined with travel time utility to compare the shortest route with the route selected by each model. The model-selected routes were not always the shortest routes and differed between trip purposes, showing how visual route preferences can be mapped across a pedestrian network.
Overall, this thesis shows that pedestrian route choice is shaped by more than travel time alone. Visual street-level information can improve pedestrian route choice models, but the required level of model complexity depends on the purpose of the analysis. The models in this thesis should be seen as exploratory tools for analysing stated pedestrian route preferences and visual route attractiveness, rather than direct prediction tools for actual pedestrian flows.