Research On Methods Of Video Content Analysis Based On Spatio-temporal Variation

With the development of the technology of internet and multi-media, online videos are becoming one of the main ways to get information, learn knowledge and have fun for users on the Internet. So, recently, video content analysis is a key problem for researchers in the field of computer vision. Video content analysis algorithm is developed from image content analysis, aiming at the automatically analyze input videos, so as to understand the related semantics for video classification and retrieval. Video content analysis methods have a broad application background among fields like medicine, safe control, and military technology.This paper focuses on the methods of video content analysis and proposes certain amount of innovation in feature extraction, data label, and analysis structure. To begin with, a novel method to extract the feature of video visually, which is called Spatio-temporal Variation Histogram Descriptor, so as to recognize the content and classify the video efficiently and precisely is presented in the paper. The concept of spatio-temporal slices is introduced firstly, and then a proper grid of key lines is used to obtain the spatio-temporal slices in the spatio-temporal volume. Simultaneously, detail of this approach to generate the spatio-temporal variation descriptor on the spatio-temporal slices is defined. This method is based on the pattern analysis of spatio-temporal slices by units of shots from the whole videos, which is independent of key frames and interesting points. Experiments show that this proposed feature can describe the inner relationship of video in time and space domain comprehensively, and consequently improve the precision of action recognition and classification. Compared with other algorithms, this method has a good performance both in time cost and classification precision. Moreover, when combined with audio features, this algorithm can achieve an ideal result for the classification of videos with common semantics.While for data label, this paper applies semi-supervised learning method to realize the web video dataset classification. A modified method to standard co-training method is proposed, called pervasive co-training method. This method works well when the dataset is increasing dynamically. At the same time, in the aspect of analysis structure, this paper designs an online video classification method based on multimodal features, and thereby realizes the security supervision of the videos. This method filters the input videos by different features about audio, color motion and space-time features in a specific order. According to the definition of the illegal scenes, including horror, violence and pornography, it detects potentially illicit information in videos. Experiments show that this method can effectively improve the precision rate of detection and classification.