A Novel Predictive Modelling Approach Towards a Spatiotemporal Traffic Safety Index

Sifatul Mostafi, Asma Fariha, Mohamed El-Darieby, Khalid Elgazzar, Akramul Azim

Abstract: Traffic safety is now a major concern due to the increase in roadways, traffic, and pedestrians. An index quantifying the traffic safety of a location can help transportation authorities take necessary measures to enhance traffic safety by enforcing Intelligent Transportation System Warrants. All the existing approaches adopt descriptive, inferential, or exploratory data analysis-based statistical techniques to calculate a traffic safety index based on different spatial features of a geographic location. In this paper, we propose a novel predictive modeling approach to calculate a spatiotemporal traffic safety index incorporating both spatial and temporal features based on the National Collision Database Canada from year 1999 – 2019. We conduct Pearson chi-square feature selection and apply feature encoding with categorical embedding. We built a deep learning model based on heterogeneous spatiotemporal features to predict traffic accident severity in terms of human injury and fatality and use it as an indicator for a spatiotemporal traffic safety index. The results show that the proposed methodology can effectively generate a spatiotemporal traffic safety index based on accident severity with an accuracy of 96.26% and surpasses the state-of-the-art approach that has an accuracy of 83%.