Using Deep Learning in Severity Analysis of At-Fault Motorcycle Rider Crashes

Abstract

Motorcyclists are vulnerable highway users. Unlike passenger vehicle occupants, motorcycle riders do not have either protective structural surrounding or the advanced restraints that are mandatory safety features in cars and light trucks. Per vehicle mile traveled, motorcyclist fatalities occurred 27 times more frequently than passenger car occupant fatalities in traffic crashes. In addition, there were 4,976 motorcycle crash-related fatalities in the U.S. in 2014—more than twice the number of motorcycle rider fatalities that occurred in 1997. It shows that, in addition to current efforts, research needs to be conducted with additional resources and in newer directions. This paper investigated five years (2010–2014) of Louisiana at-fault motorcycle rider-involved crashes by using deep learning, which is a competent tool for mapping a high-multidimensional input into a smaller multidimensional output. The current study contributes to the existing injury severity modeling literature by developing a deep learning framework, named as DeepScooter, to predict motorcycle-involved crash severities. The final deep learning model can predict severity types with 100% accuracy with training data, and with 94% accuracy with test data, which is not attainable by using a statistical method or machine learning algorithm. The intensity of severities was found to be more likely associated with rider ejection, two-way roadways with no physical separation, curved aligned roadways, and weekends. It is anticipated that the DeepScooter framework and the findings will provide significant contributions to the area of motorcycle safety.