Incident detection based on semantic hierarchy composed of the spatio-temporal MRF model and statistical reasoning

Japan and EU governments aim to reduce the mortal rate from traffic accidents by 50% at the end of 2010. To achieve this goal, sufficient information about accidents need to be gathered, so investigators can evaluate for causes and how to prevent the future accidents. In this work, we develop three algorithms for automatic event detection, and provide video clip during accidents as our results. Three algorithms are (1) logical reasoning focusing on an individual behavior, (2) logical reasoning focusing on relative behavior and (3) classification with continuous variables by hyperplane. Our algorithms utilize a semantic hierarchy composed of three kinds of operators (coordinate-class, behavior class, event-class). The three operating classes provide the context of traffic events similar to the understanding of a human operator on traffic scenes. We evaluate our algorithms on actual traffic scene taken for 18 months. Our algorithms can detect more than 90%.