A study of seasonal variations of traffic collisions and their machine learning prediction model for highways in B.C.

Abstract

Traffic collisions are considered as one of the world’s major public health problems. According to the World Health Organization (WHO), about 1.3 million people die every year in traffic collisions across the world and a further 20 - 50 million are injured or disabled. Various tools/methods were developed to assess highway safety. Historically, collision frequency, collision rates, linear regression and generalized linear regression, and Bayesian modeling methods have been used as the basis for safety analysis. Research has shown that there are limitations with this approach due to the non-linear relationship between collision frequency and exposure. Traffic volume is directly related to traffic exposure, traffic exposure affects collision risk and collision risk significantly determines the probability of traffic collision occurring. The different traffic exposure has a different level of effects on collision risk, and different level of collision risk results in the different probability of collision occurring and collision severity. Collision prediction modeling (Safety Performance Function) is the recommended technique for estimating road safety in the Highway Safety Manual (HSM) by the American Association of State Highway and Transportation Officials (AASHTO). However, the prediction modeling has not taken into consideration of traffic seasonal variations, collision seasonal variations and weather impacts as the annual average daily traffic (AADT) is one of main dominant variables. Previous studies indicate that weather especially winter weather condition is significantly associated with the traffic collisions. For example, studies showed that 24% of all collisions are weather-related in United States and collision risk could increase from 50 to 100 percent during precipitation. Due to climate change, weather patterns are changing, and the frequency of extreme weather events increases, which will affect highway safety and reliability. This study synthesizes the major findings and proposed methodologies from the existing traffic safety studies. Collision risks related to weather are investigated and assessed. Traditional techniques of highway safety assessment without the consideration of seasonal variations of traffic collisions, especially winter weather condition impacts in Canada, might result in underestimating the safety risk in winter weather conditions. All predication models developed to date have not taken into consideration of traffic and collision seasonal variations, and weather condition. Machine learning (ML) is able to address non-linear relationship between traffic exposure and collision frequency, to handle multivariate data, and to improve over time in the traffic collision prediction modelling. Two major highways (Highways 16 and 97) with a total length of approximate 2,500 km within Northern Region in the Province of British Columbia, Canada, are investigated. This study has mainly focused on the seasonal variations of collisions and traffic volumes to improve the highway safety. A traffic collision prediction model integrated traffic seasonal variations and weather impacts is proposed and developed by applying machine learning (ML) techniques (neural network regression) in the study. In conclusion, the proposed model is able to predict the traffic collision seasonal variations, and to provide more accurate estimate of traffic collisions with over 90% accuracy on both rural and urban highways. The model can be used to assist in developing road safety improvement policy considering collision seasonal characteristics.