Leveraging Multisource Data for Insights into Split Failure Occurrences

Abstract: Split failures pose significant challenges for traffic signal management, impacting mobility and safety. Regarding mobility, split failures indicate severe queueing and increased travel times, contributing to higher emissions from idling vehicles. From a safety perspective, split failures are highly correlated with red-light running, often leading to severe crashes. Analyzing split failures and identifying the factors influencing their occurrence is crucial for effective traffic management. This study comprehensively investigates the impact of various factors on split failures using high-resolution event-based video data from Miovision cameras at six four-way signalized intersections in Pima County, Arizona, along with crowdsourced speed data. Traffic and signal data collected over four Wednesdays between September and October 2021, from 7:00 AM to 5:00 PM, were considered for the analysis. Additionally, to ensure a balance between data completeness and granularity for useful insights, split failures were aggregated into 15-minute intervals. Negative binomial (NB), zero-inflated NB (ZINB), and heterogeneous NB (HNB) models were developed to account for the zero-inflated and overdispersed nature of the split failures data. Study findings indicate that the ZINB model provided the best fit, likely attributed to its ability to handle data with excess zeros. The ZINB model identified that the major road indication, increased delay, and longer intersection crossing length contributed to increased split failures, while higher median speed and morning peak were associated with reduced split failure frequency in 15-minute intervals. These insights offer practitioners valuable guidance for optimizing signal operations to enhance traffic flow and safety.