Estimating Traffic Crash Counts Using Crowdsourced Data: Pilot analysis of 2017 Waze data and Police Accident Reports in Maryland

2018PDFTech ReportDOT-VNTSC-BTS-19-01OR02A2Human machine systemsTraffic patrolTraffic accidentsLearning (Artificial intelligence)RandomizationStatistical analysisData collectionSafety Data InitiativeCrowdsourced roadway dataPolice reported crashesRandom ForestMarylandJohn A. Volpe National Transportation Systems Center (U.S.)Flynn, Dan F.B.Gilmore, Michelle M.Sudderth, Erika A.John A. Volpe National Transportation Systems Center (U.S.)United States. Bureau of Transportation Statistics. Office of DirectorNTL-SAFETY AND SECURITY-Highway SafetyNTL-SAFETY AND SECURITY-AccidentsNTL-INTELLIGENT TRANSPORTATION SYSTEMS-Incident ManagementVolpe National Transportation Systems Center, Technical Reference CenterThe U.S. Department of Transportation (DOT) is leading a Safety Data Initiative (SDI), to enhance the analysis and visualization that informs safety policy decisions. In support of the SDI, Volpe applied machine learning techniques to assess the potential for crowd-sourced roadway data such as Waze alerts to serve as a reliable indicator of police-reportable crashes. Using six months of Waze alerts and police reported traffic accident data for Maryland in 2017, Volpe developed random forest models to estimate the number of police-reported crashes. The specific spatial and temporal patterns of the estimated crashes from the models is close to the observed police reported crashes, but not exact. The model underestimates crashes during early morning hours, and overestimates at commuting times, when the volume of Waze data is highest. The Waze crash models appear to capture unreported crashes, including minor crashes which might not require a police presence, but can seriously impact congestion. Near real time estimates of police-reported crashes using crowd-sourced traffic data such as Waze could offer an early indicator of traffic crash risk. In the next phase of the project, the team will work with state and local partners to implement case studies demonstrating specific applications of the Waze crash estimation models