Predicting pedestrian flow along city streets: A comparison of route choice estimation approaches in downtown San Francisco

Abstract

Street attributes are thought to play an important role in influencing pedestrian route choices. Faced with alternatives, pedestrians have been observed to choose faster, safer, more comfortable, more interesting, or more beautiful routes. Literature on pedestrian route choice has provided methods for assessing the likelihood of such options using discrete choice models. However, route choice estimation, which is data intensive and computationally challenging, remains infrequently deployed in planning mobility analysis practice. Even when coefficients from previous studies are available, operationalizing them in foot-traffic predictions has been rare due to uncertainty involved in the transferability of behavioral effects from one context to another, as well as computational challenges of predicting route choice with custom attributes. This paper explores a simpler method of route choice prediction, implemented in the Urban Network Analysis toolbox, which assigns probabilities to available route options based on distance alone. We compare the accuracy of distance-weighted approaches to the more detailed utility-weighted approach using a large dataset of observed GPS pedestrian traces that include numerous trips between same intersections pairs in downtown San Francisco as a benchmark. Even though a utility-weighted model matches observed pedestrian flows most accurately, a distance-weighted model is only marginally inferior, on average. However, shortest-distance and highest-utility route predictions are both significantly inferior to the utility-weighted and distance-weighted sample-enumeration methods. Our findings suggest that simplified assumptions can be used to predict pedestrian flow in practice with existing software, opening pedestrian flow predictions to a wider range of planning and transportation applications.

Keywords:

• Pedestrian path choice
• prediction
• comparative analysis
• sample enumeration
• urban network analysis
• betweenness

Notes

1 https://catalog.data.gov/dataset/tiger-line-shapefile-2014-county-san-francisco-county-ca-all-roads-county-based-shapefile

2 Sidewalk networks are often represented by multiple lines (i.e. wide sidewalks have two parallel lines with perpendicular connectors, plazas and parks have radiating networks).

3 The filtering brought the initial raw route count of 120,045 unique trajectories in the city down to 26,344 trajectories.

4 Discrete choice models commonly assume that the random component of utility is IID and takes a Gumbel distribution. The Gumbel distribution is a skewed distribution with two parameters, a location μ and scale σ. If we fix σ = 1, this helpful property allows us to formulate the probability of each choice with only the values for each μj, a multivariate generalization of the inverse logit function. See Train (2002, Chapter 3, p. 378) for the derivation.

5 We leave out highway exposure and public art variables from Sevtsuk, Basu, et al. (2020) that were found to be insignificant.

6 These coefficients do not include variables describing the number of road crossings or signalized delays, which could also impact route choice and would be desirable to add in future work.

7 Pseudo R² = 1 − RMSE(prediction)/RMSE(random).

8 This can be readily performed with the “Betweenness” tool in the Urban Network Analysis toolbox software plugin for Rhinoceros 3D (Sevtsuk, 2018b).