Vol 15, No 4 (2024)

Methods of fast synchronization of interaction of unordered active distributed digital objects grouped into clusters are proposed. Unordered objects (which do not use addresses or other individual attributes) replace addressing individual objects by addresses with addressing clusters of objects specifying sets of attributes common to such objects. This allows any number of objects to simultaneously exchange signals and messages with any number of objects. Transfer of objects into synchronous state is performed in a single-cycle manner, by sending a single synchronization signal by the objects, and then maintained for a specified time. The main type of communication between objects is wireless channels, with the transmission of optical or radio signals. Violation of synchronization is restored by the objects by their own actions. Objects can be stationary and mobile. The main features of the structure of object connections are simultaneous visibility of the state of all objects by any object of the group and simultaneous delivery of messages from a group of objects to the input of any object with bit synchronization.

The optimal control problem for one of the variants of a quasilinear system is considered. To solve it, the idea of Professor V. I. Gurman is used, who proposed to combine two variants of the expansion principle. One of them is the traditional Krotov approach, and the second is the penalty function method. The selected class of systems allows for an analytical study of the Krotov Lagrangian, which in turn leads to the formulation of the algorithm. The resulting algorithm is tested on two illustrative examples, for which minimizing sequences are constructed. The complexity of the calculations is comparable with methods based on the traditional expansion principle. The calculation results are illustrated by tables and graphs.

The article proposes a method for neural network classification of short videos. The classification problem is considered from the point of view of reducing the number of operations required to categorize videos. The proposed solution consists of using a small number of frames (no more than 10) to perform classification using the lightest neural network architecture of the ResNet family of models. As part of the work, a proprietary training dataset was created, consisting of three classes: “animals”, “cars” and “people”. As a result, a classification accuracy of 79% was obtained, a database of classified videos was formed, and an application with GUI elements was developed for interacting with the classifier and viewing the results.

Computer vision requires video stream analysis, including extracting information from frames, detecting specific objects, and collecting data about them. After detection, tracking or following objects in the video stream is often required. Non-rigidity or shape variability hinders object analysis, complicates their detection and tracking, and worsens localization.The review considers architectures, models, methods, and algorithms used in practice for detection and tracking of non-rigid objects, and highlights promising solutions.