Light-Weight Classification of Human Actions in Video with Skeleton-Based Features

Abstract

An approach to human action classification in videos is presented, based on knowledge-aware initial features extracted from human skeleton data and on further processing by convolutional networks. The proposed smart tracking of skeleton joints, approximation of missing joints and normalization of skeleton data are important steps of feature extraction. Three neural network models—based on LSTM, Transformer and CNN—are developed and experimentally verified. The models are trained and tested on the well-known NTU-RGB+D (Shahroudy et al., 2016) dataset in the cross-view mode. The obtained results show a competitive performance with other SOTA methods and verify the efficiency of proposed feature engineering. The network has a five times lower number of trainable parameters than other proposed methods to reach nearly similar performance and twenty times lower number than the currently best performing solutions. Thanks to the lightness of the classifier, the solution only requires relatively small computational resources.