Developing accident frequency prediction models for urban roads: A case study in São Paulo, Brazil

Abstract

The growing number of vehicles and the evolving behaviour of road users present new and additional challenges to road safety. Study on the variables that influence the frequency of crash occurrences such as road geometry, junction, speed and land use are needed as they have proven effects on the number and severity of crashes. In this paper, we identify and assess the variables, namely road geometry, vehicle speed, traffic volume, land use and junction type, and develop accident frequency prediction models for a main urban transport corridor in São Paulo, Brazil. Crash data was provided by the traffic management company of the city, other datasets were obtained from a mix of primary and secondary sources including roadside cameras, Geographic Information Systems (GIS) and digital mapping tools. The studied road was segmented and the coefficients associated with variables in the segments were obtained using Poisson regression through a stepwise variable selection procedure. Two models with junctions density per type (access/km, T-junction unsignalised/km, T-junction signalised/km and crossroads/km) and junction density per merged type (signalised/km and unsignalised/km) along with land use per type (commercial and residential) are developed. The junction density and land use are found to be significant and positively correlated with crash frequency. The models were evaluated by statistical means for their accuracy of predicting the crashes, and validated with additional information obtained from field observation.