Statewide Estimation of Crash-Induced Delay and Queueing by Fusing Crowdsourced and Police Data

Abstract: Accurately predicting when crashes will occur is a persistent challenge due to their unpredictable nature. As a result, a key strategy to enhance roadway safety focuses on mitigating the impact of crashes when they do happen. Many transportation agencies have adopted this approach through incident management programs aimed at reducing congestion and preventing secondary crashes. These programs require rapid, efficient responses that depend on timely and accurate information, such as reliable estimates of crash-induced congestion. This study proposes a machine learning-based method for estimating crash-induced delays and traffic congestion queue lengths by integrating multiple data sources. The analysis uses police-reported crash data and crowdsourced Waze data collected from all major thruways in New York State. A Long Short-Term Memory (LSTM) model was developed to align these data sources spatiotemporally and estimate crash-related delays and queue lengths across various crash types and severity levels. The model takes into account roadway characteristics, surface conditions, and weather. It achieved a root mean squared error of 0.85 minutes for delay estimation and 1.41 miles for queue length estimation. The model's performance was also compared with three traditional models—Dummy Regressor, Random Forest, and XGBoost—demonstrating superior accuracy in estimating both crash-induced delay and queue length. The study’s findings offer practical applications for roadway planning and driver navigation systems by providing accurate crash information for display on variable message signs, which can help drivers make informed decisions, such as choosing alternative routes. Additionally, this data could assist roadway agencies in preventing secondary crashes by improving real-time traffic management.