Predicting Red Light Running Violation Using Machine Learning Classifiers

Abstract

Red light Running (RLR) violation remains as an important road safety concern at urban intersections. Existing method have mostly used statistical regression-based methods to explore factors contributing to RLR. However, it is well-known that statistical methods are based on predefined associations among variables and are unable to capture latent heterogeneity. This study aims to classify and predict RLR using spatial analysis and machine learning (ML) methods. Georeferenced RLR violation data for the year 2016 was collected for the city of Luzhou, China. To identify violation hotpots, frequency-based clustering was carried out using collect event tool in ArcMap geographic information system (GIS). Prior to RLR prediction via ML, data imbalance problem was addressed using a random over-sampling technique. Two widely used ML algorithms, i.e., Random Forest (RF) and Gradient Boosted Decision Tree (GBDT), were then used for prediction and classification of RLR. The performance of these models was assessed with accuracy and Cohen's kappa. The results showed that GBDT had an overall accuracy of 96% outperformed the RF.