A Primitive-Based Traffic Scene Semantic Description Model

Abstract

In a traffic scene, many sensors are placed, and systems are implemented to monitor incidents. In this paper, we define an event ontology that allows natural representation of complex spatio-temporal events common in a traffic scene. A Primitive-Based Semantic Description method is proposed to model traffic scenes including roads, moving objects, and their interactions, making full use of the data of sensors. A traffic scene is the basis for moving objects, thus we propose an unsupervised clustering algorithm which pays attention to detail in order to deal with uncompleted trajectories for semantic region analysis. Under our event description framework, this model is used to connect three elements in traffic visual surveillance: low-level visual features, simple sub-events, and multi-agent interactions. The definition, syntax and compilation of script language on traffic scene are discussed. The event which is proposed for user’s demand can be defined and described, and abnormal events can be inferred by our description model for accident alarming. In addition, the method has great extensibility and maintainability to integrate different Intelligent Systems.