A Knowledge Engineering Approach for Complex Violence Identification in Movies

Along with the rapid increase of available multimedia data, comes the proliferation of objectionable content such as violence and pornography. We need efficient tools for automatically identifying, classifying and filtering out harmful or undesirable video content for the protection of sensitive user groups (e.g. children). In this paper we present a multimodal approach towards the identification and semantic analysis of violent content in video data. We propose a layered architecture and focus on ontological and knowledge engineering aspects of video analysis. We demonstrate the development of two ontologies defining violent hints hierarchy that low level analysis, in visual and audio modality, respectively should identify. Violence domain ontology, as a reality representation, defines higher-level semantics. Taking under consideration extracted violent hints, spatio-temporal relations and behavior patterns higher-level semantics automatic inference is possible.