Comparison of two simulation methods for testing of algorithms to detect cyclist and pedestrian accidents in naturalistic data

Naturalistic studies can potentially be used to detect accidents of vulnerable road users and thus overcome the large degree of under-reporting in the official accident records. In this study, simulated cycling and walking accidents were performed by a stunt man and with a crash test dummy to test how they differ from each other and the potential implications of using simulated accidents as an alternative to real accidents. The study consisted of simulations of common accident types for cyclists and pedestrians, such as tripping over a curb or falling of the bike after hitting an obstacle. Motion data in terms of acceleration and rotation as well as the state of the screen (turned on/off) was collected via an Android smartphone to use as indicators for the motion patterns during accidents. The results show that dummy data have a distinct peak at the moment of the fall as a result of not being able to break the fall. As opposed to this, the stuntman arranges himself in a way to reduce the impact when hitting the ground. In real accidents, motion patterns will probably lie in-between these two types.