Understanding Drivers’ Safety by Fusing Large Scale Vehicle Recorder Dataset and Heterogeneous Circumstantial Data

We present a method of analyzing the relationships between driver characteristics and driving behaviors on the basis of fusing heterogeneous datasources with large-scale vehicle recorder data. It can be used, for example, by fleet managers to classify drivers by their skill level, safety, physical/mental fatigue, aggressiveness, and so on. Previous studies relied on precise data obtained in only critical driving situations and did not consider their circumstances, such as road width and weather. In contrast, our approach takes into account not only a large-scale (over 100 fleet drivers) and long-term (one year’s worth) records of driving operations, but also their circumstances. In this study, we focused on classifying drivers by their accident history and examined the correlation between having an accident and driving behavior. Our method was able to reliably predict whether a driver had recently experienced an accident (f-measure \(=\) 72%) by taking into account both circumstantial information and velocity at the same time. This level of performance cannot be achieved using only the drivers’ demographic information or kinematic variables of operation records.