비디오 데이터의 객체 움직임 인식을 위한 의미적 이벤트 검출

Abstract

For semantic search of video data, semantic event representations of object motion are being actively studied. However, the existing studies are mostly about content-based video search using low-level features and predefining notes for the video data, and inadequate to interpret and process the conceptual high-level motion of human beings. In this paper, relationship between objects is represented by a hierarchical structure of linguistic concepts to provide semantic representations of video events, and a method that maps high-level motion information, which is for semantic shot detections of video, to low-level features of the video is suggested. In this paper, first, conceptual relationship between verbs is represented using a hierarchical structure to represent motion event of object in the video shots. Event verbs are located on top of the hierarchical structure, and verbs for each event are arrayed as instances. Second, rules of the intermediate level that links low-level features to high-level motion information are defined to detect concept-based shots from video images. For that reason, characteristic values of moving objects are extracted throughout the video data processing. Motion recognition rules suggested in this paper define primary rules for spatial relationship using locations of still objects and moving objects, and, for recognition of detailed actions, motion of each consecutive frame is recognized by distances and directions between objects. In this way, events represented by shots are extracted, and, consequently, indexing of the shots is carried out by analyzing consecutive patterns of the recognized motion. This is to allow users to use indexes when they want to search videos related to certain events. Third, when searching events, ontology is established to semantically search events which are represented differently depending on application. The lexical relationship between video shots and events, indexed for the final ontology search, is described with the standard language of the semantic web OWL/RDF(S). his is to provide semantic search and inference using characteristics of IS-A relationship and synonyms in ontology, and convenient search services for users. For semantic event detection of video data suggested in this paper, motion recognition experiments of object have been carried out in two ways. The average precision rate of motion recognition for each shot was 85% in the test with the modelling suggested and precision rate for event detection was 80%.