Abstract
The article discusses the problem of recognizing human actions based on skeletal data using convolutional neural networks (CNNs). The analysis of modern methods of Human Activity Recognition (HAR) and the advantages of skeletal data representation is carried out. The NTU RGB+D dataset was used for the experiments. A method for selecting keyframes using the K-means algorithm and subsequent optimization of the sequence of frames is proposed. Experimental results have shown that the proposed approach provides high recognition accuracy while reducing computational costs. The highest accuracy was achieved for the actions "walking" (0.96), "sitting" (0.95) and "pushing a person" (0.93). It has been found that the greatest difficulties arise in recognizing gestures associated with hand and head movements due to the similarity of the spatial and temporal characteristics of the skeleton. The results obtained confirm the effectiveness of using sequences of keyframes and convolutional neural networks for tasks of recognizing human actions in video streams.