Studying the effects of weather and roadway geometrics on daily accident occurrence using a multilayer perceptron model

One of the most common, yet dangerous, events that people face each day is driving. From unpredictable weather to hazardous roadways, there is a seemingly endless number of factors at play that can lead to vehicular accidents. Therefore, attempting to predict these accidents is a timely topic in today's research spectrum. The data used in this research consists of historical accident records from Hamilton County, Tennessee beginning in 2016 and continues to be updated daily, as well as the associated weather occurrences and roadway geometrics. To enhance heterogeneity a procedure was performed that generated non-accident traffic data based on our actual traffic accident data. This procedure is called negative sampling. These different data sets were combined and placed through a Multilayer Perceptron (MLP) machine learning model. The end results displayed a high collective correlation between accident occurrence and the various features considered in our proposed model, allowing us to predict with 77.5% accuracy where and when an accident will occur.