In computer vision, identification of actions of an object is considered as a complex and relevant task. When solving the problem, one requires information on the position of key points of the object. Training models that determine the position of key points requires a large amount of data, including information on the position of these key points. Due to the lack of data for training, the paper provides a method for obtaining additional data for training, as well as an algorithm that allows highly accurate recognition of animal actions based on a small number of data. The achieved accuracy of determining the key points positions within a test sample is 92%. Positions of the key points define the action of the object. Various approaches to classifying actions by key points are compared. The accuracy of identifying the action of the object in the image reaches 72.9 %.