Institute of Transportation, MOTC

       To improve the highway safety, the regulation and management of driving behaviors is one of the most critical issues. The highway safety management policy which made by the concerned departments usually depends on the historical traffic crash events or macro traffic flow. However, it is hard to have a depth knowledge of microscopic driving behaviors due to the limited data granularity, which is insufficient to achieve the safety regulation improvement. Related studies usually take near crash events extracted from the driving behavior data as the main input of crash prediction and safety improvement. Nevertheless, the driving behaviors varies from drivers, not all drivers react dangerous while emergency. Different from near crash event which owns static and fixed standard, the standard of abnormal driving behavior is dynamic and comparatively. This research uses Variational Autoencoder to detect the abnormal driving behaviors from the data of highway bus and explores the relations between the historical traffic crashes with the comparison. This research can discover the potential risk before crash, which makes highway safer actively.