Methods for Estimating Bicycle Volumes at Urban Intersections and Road Segments

Abstract

Accurate estimation of bicycle volumes is essential for transportation planning, safety analysis, and infrastructure design, particularly for evaluating risks faced by vulnerable road users. Bicycle exposure is commonly expressed as annual average daily bicycle (AADB) volume; however, many jurisdictions lack sufficient continuous or short-term count data to estimate AADB reliably across their transportation networks. This limitation restricts the ability of agencies to conduct comprehensive safety analyses and to make data-driven decisions regarding bicycle infrastructure investments. Consequently, there is a need for practical methods to estimate bicycle volumes under limited data availability. This thesis investigates several approaches for estimating bicycle volumes at intersections and road segments when observed count data are limited or unavailable. The research is structured as four complementary studies. The first study evaluates the spatial transferability of existing direct-demand (DD) models developed in different jurisdictions. The results demonstrate that naïvely applying these models to new jurisdictions can result in substantial estimation errors, highlighting limitations in their direct applicability. The second study examines methods to improve transferability of DD models through local calibration. The findings show that even a small number of locally observed count sites can significantly improve model performance, with regression-based calibration approaches providing notable reductions in estimation error. The third study explores the use of large language models (LLMs) as an alternative approach for estimating bicycle volume levels at intersections. The results indicate that LLMs can provide reasonable estimates of bicycle activity categories using limited input information, suggesting that they may offer a practical and accessible tool for jurisdictions lacking extensive data or modeling expertise. The fourth study investigates road segment-level bicycle volume estimation using a range of modeling approaches, including traditional regression models, machine learning techniques, LLM-based methods, and spatial regression models. The results demonstrate that incorporating spatial relationships between network elements improves estimation accuracy, emphasizing the importance of spatial dependence in modeling bicycle activity. The findings of this thesis contribute to the development and evaluation of practical methodologies for estimating bicycle exposure under limited data conditions. The results provide guidance for transportation agencies on selecting appropriate modeling approaches and highlight the potential of combining calibration, artificial intelligence, and spatial modeling techniques to improve bicycle volume estimation. These advancements support more reliable safety analyses and more informed planning decisions aimed at improving conditions for vulnerable road users.