The paper considers the problem solution of classifying the type of physical activity of a person according to visual data. The authors propose using of deep neural networks to determine the type of activity. The recognizing human activity from video data or a single image systems are currently actively used in various areas of human activity. As the example we can take the system for monitoring the effectiveness of enterprise employees. So solving the problem of recognizing human actions from visual data is an actual task. The authors developed an algorithm for determining the physical activity type by visual data based on the DenseNet121 and MobileNetV2 models. Then the deep neural network model was built and hyperparameters were selected, because pre-trained networks did not provide the required accuracy of detecting the type of physical activity. The software implementation of the model is made in the IDLE environment in the Python programming language. Experimental studies performed on a specialized UCF50 dataset containing 50 different types of human actions confirm the effectiveness of using the proposed approach to solve the problem. Additionally, the representativeness of the test data set was increased with the help of video sequences obtained from YouTube. Purpose – development of an algorithm for determining a person’s physical activity based on visual data. Methodology: in the work the methods of computer vision, deep learning methods and object-oriented programming methods were used. Results: an algorithm for tracking a person’s physical activity based on visual data using deep learning technologies has been developed. Practical implications: the obtained results can be used in human activity monitoring systems, for example, in tracking criminal activity, in medical diagnostics, in tracking the activity of office employees, etc.

Sign recognition is an important task, in particular for the communication of the deaf and hard of hearing population with people who do not know sign language. Russian sign language is poorly studied, Russian sign language of the Siberian region has significant differences from others within the Russian language group. There is no generally accepted data set for Russian Sign Language. The paper presents a gesture recognition algorithm based on video data. The gesture recognition algorithm is based on the identification of key features of the hands and posture of a person. Gestures were classified using the LSTM recurrent neural network. To train and test the results of gesture recognition, we independently developed a data set consisting of 10 sign words. The selection of words for the data set was made among the most popular words of the Russian language, as well as taking into account the maximum difference in the pronunciation of gestures of the language dialect of the Siberian region. The implementation of the gesture recognition algorithm was carried out using Keras neural network design and deep learning technologies, the OpenCV computer vision library, the MediaPipe machine learning framework, and other auxiliary libraries. Experimental studies conducted on 300 video sequences confirm the effectiveness of the proposed algorithm.