Estimating annual average daily pedestrian traffic volume at intersections

Abstract

Pedestrian exposure at intersections is a critical input for jurisdictions developing pedestrian-centric strategies and is typically quantified as the annual average daily pedestrian traffic (AADPT). The ability to estimate AADPT across all intersections depends on the availability of pedestrian volume data – specifically, the number of sites with continuous count (CC) stations, sites with only short-term counts (STCs), and sites with no available data. At sites with CC stations, AADPT can be directly calculated. Where only STCs are available, expansion factors derived from CC stations are used to expand STCs to AADPT. For sites lacking pedestrian volume data, Direct-Demand (DD) models are commonly employed to estimate pedestrian exposure based on land use, socioeconomic, and transportation attributes. This paper-based PhD thesis addresses gaps in the existing literature by exploring methods for estimating pedestrian exposure at intersections under varying data availability scenarios. Objectives 1 and 2 focus on estimating pedestrian exposure at sites with no available data in jurisdictions that lack sufficient sites to locally develop DD models. Objective 1 examines the direct (or naïve) spatial transferability of DD models developed in other jurisdictions. Performance was found to be inconsistent, varying based on the similarity between the characteristics of the jurisdiction where the model was developed and where it was applied. In Objective 2, it is assumed that a limited number of CC sites are available within the study jurisdiction, enabling the adjustment (or local calibration) of transferred DD models. This approach improved the transferability of DD models, achieving performance levels comparable to those of locally developed models. Continuing on jurisdictions that lack sufficient data to develop their own DD model, Objective 3 proposes a novel method to estimate pedestrian exposure leveraging AI Large Language Models (such as ChatGPT), satellite imagery, and pedestrian volume data from a few key sites. The performance of this method rivaled – and in some cases surpassed – existing conventional methods like DD models, all without requiring complex statistical models or extensive datasets. Objective 4 shifts the focus to sites where STCs are available, providing a detailed application of the expansion factor method to expand 8-hour STCs. Specifically, it examines grouping sites with CC stations (where expansion factors are derived) into factor groups based on similar activity patterns. To achieve this, temporal indicators capturing different types of seasonality were developed, and models were built to associate STC sites with an appropriate factor group. The results indicated only marginal improvements when using the factor group approach compared to applying a single factor group (i.e., the average across all sites). Objective 5 examines the scenario in which CC stations are unavailable within a jurisdiction, preventing the local calculation of expansion factors. The spatial transferability of expansion factors across jurisdictions is investigated. The findings indicate that transferring expansion factors across jurisdictions with similar characteristics in terms of weather and school holiday periods proved to be of practical value, with only minor performance degradation compared to expansion factors developed within the local jurisdiction. Objective 6 evaluates various methods for estimating pedestrian exposure under different data availability scenarios. The performance of these methods was analyzed based on the number of sites with available data, resulting in practical guidelines for practitioners. Specifically, it provides recommendations on the minimum number of sites with pedestrian volume data needed for the local development of DD models, compared to alternative approaches such as spatial transferability and local calibration of DD models. Objective 7 introduces a novel approach to disaggregating pedestrian volumes from the intersection level to the crosswalk level. Land use models were developed and compared to methods that incorporate crosswalk-specific volume data from STCs. Overall, methods using STC data outperformed the land use models.