A neural network based algorithm for classification of sets of human body keypoints

Human action recognition has attracted a lot of attention due to rapid development of surveillance systems and intelligent medical applications. There are multiple approaches to action recognition, such as processing images as a whole or performing analysis of a specific set of features. One of the prominent approaches is to use pose estimation to perform posture analysis. In this paper we consider the latter and present a neural network based algorithm for classification of sets of human body keypoints. The key feature of the algorithm is data preprocessing approach, which allows to increase generalization of the neural network performing classification. We also present a new dataset for the task of human body keypoints classification. The algorithm was evaluated on the presented dataset achieving 89.6% accuracy. The dataset will be available at https://github.com/oKatanaaa/SSAU-Human-Pose-Classification.