Description: At present, the information technologies development and the modern information society behavior has led to the huge volumes emergence of primarily multimedia information, as the most convenient point of view by human perception. At the same time, the necessary data search becomes more and more difficult, especially if there is a need to search for the content of multimedia data, especially video where exist approach so-called CBVR (Content-Based Video Retrieval). At the same time, there do exist many problems. First of all there is always a time limit for finding the necessary information, but the volumes processed information are not just large, but are constantly increasing. One way to obtain the necessary information is a segmentation of the raw data into homogeneous segments with the subsequent possible replacement of the corresponding segment with key frames, which will substantially reduce the information needed for processing during content-based search in video data. Given that the video can be represented as a set of frames, it is possible to use a time series analysis. In this paper, we consider the problem of fast properties change detections in multidimensional time series using an ensemble of matrix autoregressive (MAR) models. The use of the proposed approach to the formation of a model ensemble makes it possible to adjust each adaptive MAR model using its own identification criterion with different depth of memory. The proposed approach is intended for use in clustering-segmentation problems of high-margin multidimensional data, such as video. The constructed models ensemble is simple in terms of software implementation and allows you to quickly process video data online.
Keywords: multidimensional time series, video data, ensemble of models