Accident Hotspot Prediction and Prevention Using Machine Learning

Abstract

In cities the rising rate of road accidents is one of the major threats to not only the safety of the citizens but also the overall effectiveness of the traffic system. Conventional reactive measures like post-incident review and blanket safety initiatives tend to be deficient in keeping accidents at bay prior to the crash. In an attempt to fill this gap, the present research offers a proactive data-driven approach built upon machine learning (ML) methods that aim at predicting the sites of accidents (the so-called hotspots). Historical crash reports, environmental factors (e.g., weather, lighting), time series (e.g., peak times, time of the year), and structural characteristics of road networks will be measured expressly to analyze their relations with the destination of the crashes. Traffic incident logs, weather archives, and geospatial road data will all be publicly available datasets which will be used to train and validate such ML models as Random Forests, Support Vector Machines (SVM), and Neural Networks. The evaluation of these will be on the measures of their prediction of high-risk areas in terms of their accuracy, precision and recall. The system can therefore facilitate the smooth running of traffic by allowing prompt responses before the incidence of accidents occur to assist the traffic authorities to maximize the patrol units, enact local safety measures, upgrade urban infrastructure, and eventually, minimize accidents in the metropolitan road systems.