Analysis of dangerous driving behavior from driving log data for traffic accident risk prediction

Abstract

Since traffic accident is a problem that causes socioeconomic issues in the nations, many endeavors have been done to lessen or prevent traffic accidents. In this study, to reduce future traffic accidents by using Digital Tachograph (DTG) data and deep neural networks, we first study the correlation between dangerous driving behavior and historical accident records and then build a deep learning model to predict traffic accident risk by utilizing the knowledge from the previous step's findings. Our findings reveal that dangerous driving behavior highly correlates to traffic accidents concerning the locations at the district level around 0.71, while 0.90 when do not consider Overspeed behavior. In addition, for both geographical and temporal correlation, Sudden U-Turn (0.83), Rapid Acceleration (0.83), and Sharp Turn (0.81) are the top three dangerous driving behavior that has the highest average correlation to the traffic crashes. Besides, our proposed deep learning model outperforms the baselines with the average MAE and RMSE scores of 3.302 and 5.335 in the prediction phase. It shows that we can improve the prediction performance by up to 14% by using the dangerous driving behavior with our proposed deep learning architecture. We expect that the proposed model will help reduce future traffic accidents by integrating it into an existing intelligent traffic system.