Analysis and prediction of traffic accidents at urban intersections

Abstract

Traffic accidents generate significant harm, injuries, and fatalities on a global scale, making their investigation an important topic of study. In the initial phase of the study, we examine the influence of three micro traffic parameters, namely standstill distance, headway duration, and following distance oscillation, on both traffic flow and intersection safety. A data-driven approach is proposed for identifying the Pareto optimal sets of parameter combinations that both maximise the flow volume and minimise the risk of crashes. The analysis of traffic flow is conducted by considering different scenarios involving standstill distances ranging from 0 to 5 metres, headway times of 0.5 seconds, 0.9 seconds, and 5 seconds, as well as following distances spanning from 1 to 10 metres. The trajectories derived from the microsimulation model are further examined using the surrogate safety assessment model to ascertain the distribution of time-to-conflict between vehicles. This analysis facilitates the estimation of risk for crash by employing extreme value theory. The findings from the case study of Lewvan Drive and 13th Avenue intersection suggest that utilisation of headway times of either 0.5 or 0.9 seconds, in conjunction with standstill distances exceeding 2 metres and following distance fluctuations ranging from 1 to 9 metres, guarantees the mitigation of crash risks to a minimum level, while simultaneously resulting in maximum traffic flows. The subsequent phase of the research endeavours to construct dynamic forecasts of accident rates at intersections by taking macro traffic variables into account. Additionally, it evaluates the predictive efficacy of statistical models, machine learning methods, and neural network algorithms. The initial step involves conducting Pearson's correlation and statistical analysis to ascertain the associations between the macro variables, namely the number of accidents, average daily traffic on weekdays for major and minor roads, number of legs at the intersection, traffic signal conditions, intersection location, and peak hours. A threshold value of 0.7 is employed to verify the existence of collinearity among the variables. Based on the correlation analyses, the study further employs prediction models of three streams, including statistical analysis (Negative Binomial Model), machine learning algorithm (ARIMA Model), and neural network (Multi-Layer Perceptron Model). All models leverage a common dataset, which is transformed into an hourly time series prior to the application of the models. The results of this study underscore the progress made by various prediction algorithms in accurately anticipating the incidence of traffic accidents at crossings, as well as the interrelationship between traffic volume and signal characteristics. The study's findings offer valuable insights and a framework for policymakers to effectively implement safety regulations and ensure satisfactory traffic flow. Additionally, these findings can aid in reducing accidents and optimising roadway capacity. Furthermore, the study provides valuable insights for drivers, helping them understand the importance of maintaining safe distances while driving and ultimately reducing risky manoeuvres while maximising traffic flow. Subsequently, the utilisation of the dynamic accident prediction model empowers policymakers to assess and forecast the efficacy of a safety measure with regards to factors such as traffic volume and intersection control through a comparative analysis of the changes in the variables both before and subsequent to the implementation of the intervention.