Research On Video Structure Method Based On Content

With the development of information and technology, there is an enormous increment of the video. As the cost by manual work is too high and the efficiency is too low, an automatic and intelligent method is necessary to analyze the video content. In particular for the surveillance video which contains a large number of useless images, there are many problems such as the huge redundancy, fast review of key information, classification of the video content and so on. Therefore, it's of great importance to extract the useful content of the video and analyze the video structure.The video structure method in this thesis contain video concentration in time domain, key frame extracting, feature extracting and scene classification. The main contents include:(1) To deal with the problem of the large video, a method of video concentration is proposed. The dynamic frames can be detected by the method of frames subtraction, Gaussian background modeling and visual background extractor(ViBe) respectively. Results of different methods are compared and analyzed by applying to different surveilance videos. We found that The ViBe method has a faster speed than the other two methods. Besides, on the base of not increasing more inspection rate, the rate of missed diagnosis reduce greatly and video redundancy rate is below 10% on average, which greatly improve the efficiency of video processing.(2) For the problem of the lack of video information structure and fast preview, the algorithm of key frame extraction is researched. For various video scenes, different algorithms are adopted. For indoor scene video, algorithm based on histogram comparison between frames is studied. Based on that, the histogram comparison between frames based on partitioning and multi- threshold is studied and the accuracy of this algorithm is proved. As for vehicle monitoring video, this paper mainly focused on virtual loop method. The accuracy of this algorithm used in vehicle monitoring video is proved. For the vehicle monitoring video with no traffic jam, the average accuracy would be reaching 95% or higher. What's more, as for the key frame extraction, the influence of the speed of moving object is analyzed. The result shows that the faster the speed is, the greater the ratio of key frames extraction will be.(3) Aiming at the problem of the video scene diversity and target classification, features are extracted with regard to the foreground object. Meanwhile, we studied the random sample consensus(RANSAC) a lgorithm based on SURF feature, by which the content of the key frames is matched and the classification of video scene is realized. Compared with the general video scene classification method, this algorithm show a better accuracy.In conclusion, in this article, by the study of the above methods, the video structure analysis of the video is realized. The different monitoring video structure analysis method is studied and summarized, which has a great significance for the follow-up research work.