Automated statewide estimation of crash-induced delay and queueing using crowdsourced data

Abstract

Due to the unpredictable nature of crashes, accurately predicting when crashes will happen is challenging. Therefore, a key strategy for enhancing safety focuses on mitigating the impact of crashes when they do occur. Many agencies have adopted this approach by implementing incident management programs designed to reduce congestion and prevent secondary crashes. These programs require quick and efficient responses, which depend on timely and relevant information, such as accurate estimates of crash-induced congestion. This study introduces a method for estimating crash-induced delay and traffic congestion queue length, using machine learning and fusing multiple data sources. Police-reported crash data and Waze crowdsourced data were collected for all thruways in New York State. A hybrid model combining XGBoost, autoencoders, and gated residual networks was trained using the spatiotemporal alignment of multiple data sources. This model enables statewide estimations across different crash types and severity levels, accounting for roadway, surface, and weather conditions. The proposed model achieved an average error of 0.628 min for estimating crash-induced delays and 0.768 miles for queue length estimation. Its performance was evaluated against six state-of-the-art benchmark models, and the results demonstrated that our model consistently outperformed all others in both delay and queue length predictions. These findings have practical implications for roadway planning and traffic management, particularly in enhancing driver navigation by providing accurate crash-related information through variable message signs. This information could help drivers make informed route choices, including potential detours, while also providing valuable data to roadway agencies to prevent secondary crashes.