At the moment, a module for determining objects (people) from the streaming video data of CCTV cameras has been developed, thus the first three subtasks of the problem have been solved. For this purpose, the YOLO architecture was used, followed by additional training using additionally collected and processed data sets in the amount of 1000 images. As a result, we managed to improve the model for determining people for the environment we need (premises, hangars, workplaces, etc.). In particular, we conducted a search and analysis of alternative works describing models and methods of image processing, finding objects and classifying actions. As a result, the YOLO7 algorithm with a high rate in finding objects was chosen as a model and algorithm for finding a person in the frame. To classify actions at a given point in time, we chose an algorithm for extracting a skeleton from an image of a person or people, analyzing the position of the limbs and further classifying actions based on these data. The latter approach greatly complicates the task, but this method allows you to create large datasets from various sources and domains. Examples of finding objects are shown in Figures 2 and 3.
