Modification of software tool for human activity classification

Siberian Federal University, 79 Svobodny pr., Krasnoyarsk, 660041, Russian Federation

^* Corresponding author: vladmymlikov@mail.ru

Abstract

The work is devoted to the modification and improvement of a neural network system for classifying human actions by changing the position of points of his skeleton. Within the framework of this paper, the description and results of some experiments on modification of the system developed by the authors several years ago will be presented. The experiments include new methods of augmenting the training data, removing some elements of the network, using a completely different representation of the data as a sequence of pictures rather than a set of points, using memory values and the state of the hidden LSTM cell as the output of the network instead of its direct output, and a variant of the network with the introduction of a new layer of feature encoding at the stage of sending skeletal point data to the recurrent LSTM network. The network with different layer configurations was trained multiple times, and the run data was averaged to remove the influence of randomness. In the end, the best network for the network was the one with pre-coded features.