Maneuver-based Driving Behavior Classification Based on Random Forest

Driving behavior classification is highly correlated with vehicle accidents and injury. Automatically recognizing different driving behaviors is important for improving road safety. This article proposes a maneuver-based driving behavior classification system. For each driving maneuver, we first generate driving behavior events based on its given timestamp using three different strategies. Then, 19 temporal features of each behavior event are calculated using signals captured by accelerometers, gyroscopes, and GPS. Next, reliefF is incorporated for selecting features. Finally, random forest is used for classifying maneuver-based driving behaviors. Experimental results using the UAH-DirveSet show that our proposed system can achieve an averaged F1-score of 70.47% using leave-one-driver-out validation. For different maneuvers, we find that the highest F1-score is obtained for braking which is 75.38%.