Graph Construction Based on Temporal Information and Frame Sampling for Human Action Recognition

With the innovation and progress of data analysis, human action recognition has become a significant research direction with broad applications in many situations. We propose a skeleton temporal graph (STG) based on graph signal processing (GSP), graph spectral domain, and human action recognition. The temporal information between adjacent frames of action data is extracted by uniform sampling and redefining temporal edge weights. We reconstruct the graph Laplacian matrix from the skeleton and temporal information. According to the graph Laplacian matrix, the coefficient matrix used for classification is calculated by spectral graph wavelet transform (SGWT). In addition, we use a recent classification method eXtreme Gradient Boosting (XGBoost), to improve experimental accuracy. Our method outperforms the existing approach when applied to three publicly available datasets.