The purpose of the work is to increase the accuracy of finding time stamps of scene changes in a videostream or videorecording by developing methods and algorithms, as well as verifying them in practice, with the aim of the following application of various transformations to individual scenes of a video recording or their independent analysis. The basis of the developed methods are adaptive approaches based on machine learning, which were tested on various available datasets, which proved the high efficiency of the proposed methods. In the field of video processing, scene boundaries localization is one of the important tasks. Known methods either do not solve the task accurately enough, or become too complex and demanding on computing resources, but still do not achieve high quality indicators.
Splitting the whole video into scenes allows to: divide the entire video stream into relatively separate and useful parts (for various reasons: advertising, effective pause); analyze the presence of specific characters in the observed scene and their behavior
(what exactly the actors are doing; occurrence frequency analysis – this can be useful for security reasons, to find episodes of video recordings with a certain activity; and for broadcasts of competitions, to find a specific or culmination moment, or track
player actions and perform other game analysis); for individual scenes further annotation, determine their characteristics; track people and their actions or the movement of objects with a wide range of applications, as well as detect important moments in long videos or video streams. Video scenes localization allows faster and more convenient navigation on any video recording: a video film, a video recording of a show, or, say, a recording of a game (in particular, a computer one) – to find the moments of the greatest interest; and many other applications. Also, the division into scenes will undoubtedly improve the user experience (viewing convenience), because
instead of moving in time (backward and forward rewinding, jumping for a certain time, when it is difficult to localize an event in the record of interest), there will be an opportunity to quickly navigate between individual scenes, as, for example, in a YouTube video labeled with time ticks.