Research On Video Content Analysis Technology And Application Based On MPEG-2

Video Content Analysis(VCA) has been one of the most popular research in multimedia field, which has obtained some achievements and has been widely used in video copy detection, in-video advertising and video retrieval, etc. The key of video content analysis is how to extract the video feature that can make the video content be descripted effectively and comprehensively. In recent years, the content image analysis and processing technology, image local features, sparse theory, etc. have been continuously developed. Therefore, these new theories and methods are introduced to the video content analysis in this paper, and further study of video feature extraction and characterization is made, then the features are applied to near-duplicate video detection and content-based video ads insertion. The main contents include the following sections:1. A spatial and temporal features extraction and representation method based on MPEG-2 video is proposed.Due to the lack of feature representation in existing video content analysis method, spatial and temporal features in MPEG-2 videos are extracted according to the characteristics of video coding format to represent video content comprehensively. Firstly, video key frames are obtained based on visual saliency model, then the spatial features, HSV histograms and ORB features(Oriented FAST and Rotated BRIEF) are extracted from key frames, and ORB features are made into sparse representation, at the same time, motion vectors(MV) from the video bitstreams are exploited to build MV angle histograms as video temporal features. Thus, multi-dimensional video features combined with spatial and temporal features are obtained to represent video content.2. A video near-duplicate detection method based on MPEG-2 spatial and temporal features is designed.In this paper, the extracted video spatial and temporal features are applied to video near-duplicate detection. With the similarity of each feature being compared respectively and the decision fusion method based on voting method being used, the content similarity between query video and reference video is obtained, which near-duplicate copy detection judgement are presented. The experimental results on benchmark datasets show that the extracted features in this paper can resist various copy transformations, and the proposed near-duplicate detection method presents better accuracy and detection speed.3. A content-based video ads insertion method is designed.Ads are inserted into the target video on a fixed time point in the current video ads insertion method, which causes the serious interference when video is playing and results in viewers’ resistance to advertised goods. Therefore, the existing content-based ads insertion methods have been improved in this paper. According to video temporal and spatial features as well as video structure, the content similarity between ads and target video is calculated to obtain a proper insertion position, then ads insertion based on video content are realized. In this paper, the subjective evaluation experimental results show that the proposed method has less interference to viewers, compared to the fixed-point insertion method.