Recognizing Potential Traffic Risks through Logic-based Deep Scene Understanding

Automatic recognition of risks in traffic scenes is a core technology of Intelligent Transportation Systems. While most of the existing work on traffic risk recognition has been done in the context of motion prediction of vehicles relying on directly observed information (e.g., velocity), exploiting implicit information inferable from observed information (e.g., the intention of pedestrians) has rarely been explored. This paper proposes a novel risk prediction model that uses abductive reasoning to infer implicit information from observations and jointly identifies the most-likely risks in traffic scenes. To evaluate the authors model, create a novel benchmark dataset that contains a wide variety of risk prediction problems. The authors experiments indicate that the abduction-based framework has a great potential for solving risk prediction problems. The developed dataset is made publicly available for research purposes.