의미적 움직임 모델링 기반 비디오 내 이동 객체 그룹화

Abstract

Due to the rapid spread of the web, It has increased interest in multimedia content. until now, study of video retrieval was based on the low-level elements like color, texture, shape. However, users prefer semantic-based video retrieval to others. because, semantic-retrieval means understanding of relation about Instances in video and analysis. This paper was focused on "How to understand event in video." One major goal of this research is to accomplish the automatic extraction of feature semantics from a motion and to provide support for semantic-based motion indexing/retrieval/management. So, I proposed model used for spatio-temporal relations. and I got a moving object information using low-level information. Analysis process consists of Static analysis, Dynamic analysis and DTW analysis. first, static analysis is understanding only one frame at time. and dynamic analysis means understanding two or more frames in movies. for example, moving object is outter from other object. In this case, static analysis result is "OUT". but considering two or more frames, we will can know relation between two instance. like "approach", "depart", "go away". but these process only can know relation about approach, enter, depart...etc. we need understanding "where to go." as well. So, I used DTW(Dynamic Time Warping) algorithm. This algorithm provide more directive comparing method between two curves. of course, from now on, we can know relation about Instance and moving direction also. This paper ends with the presentation of experiments and results showing both processes are indeed capable of recognizing simple events out of video streams. Applying The proposed system, I evaluate right_result, wrong_result, right_miss_result So, the Recall Rate has a 0.78. and Precision Rate has a 0.74. In addition, the analysis shows 88% accuracy.