Human Action Recognition Based on STDMI-HOG and STjoint Feature

Abstract

More and more attention has been focused on the human action recognition domain. The existing methods are mostly based on single-mode data. However, single-mode data lacks adequate information. So, it is necessary to propose methods based on multimode data. In this paper, we extract two kinds of features from depth videos and skeleton sequences, named STDMI-HOG and STjoint feature respectively. STDMI-HOG is extracted from a new depth feature map Spatial-Temporal Depth Motion Image by Histogram of Oriented Gradient. STjoint feature is extracted from skeleton sequences by ST-GCN extractor. Then two kinds of features are connected to make up a one-dimensional vector. Finally, SVM classifies the actions according to the feature vector. To evaluate the performance, several experiments are conducted on two public datasets: the MSR Action3D dataset and the UTD-MHAD dataset. The accuracy of our method on two datasets is compared with the existing methods, and the experiments prove the outperformance of our method.