Risk Estimation of Interactions of Right Turning Vehicles and Vulnerable Road Users

Abstract

The risk assessment approach presented in this paper aims for assisting right-turning vehicles by sending I2X messages with warnings. The approach is based on stationary long-term observations of cyclists and motorists approaching an intersection by a multi-camera system. In order to improve risk assessment of interacting cyclists and motorists it is important to predict the most probable paths of the interaction partners. On the basis on trajectories of two months the Main Traffic Flows are clustered and probable paths for each cluster are modeled. Additionally, a neural network classifies each current trajectory’s MTF for providing a confidence that a road user will take a certain path and destination. Finally, risk assessment was realized on the basis of a decision tree (DT), which adapts human expert risk assessment by supervised learning: the DT was trained by selected annotated traffic situations containing different risk levels (1. controlled encounter situation, 2. slight conflicts, 3. severe conflicts and 4. crashes). Once trained, the DT provides instant risk assessment based on the trajectories and its predictions, which can be sent to the conflicting road users by I2X messages. Though DT’s have known weaknesses like instability, they allow insights on the parameters leading to dangerous situations. Currently, these methods are running on AIM Research intersection, Brunswick, Germany, to test their appropriateness to predict critical interaction situations between right turning motorists and straight ahead driving cyclists in order to avoid severe injuries. The results shown in this paper were obtained in the EU funded project XCYCLE (Advanced measures to reduce cyclists' fatalities and increase comfort in the interaction with motorized vehicles, Grant Agreement number: 635975).