Occluded skeleton-based multi-stream model using Part-Aware Spatial–Temporal Graph Convolutional Network for human activity recognition

Abstract

Human activity recognition using skeleton data has engrossed significant research attention in pattern recognition due to its broad applications. However, occlusion remains a major challenge in activity recognition. In this paper, we propose a multi-stream part-aware occluded skeleton-based graph convolutional network designed to improve predictions in the presence of occlusions. The model consists of three key modules: the Input Inhibition Module for Skeleton Sequences, which handles incomplete or occluded skeleton data; the Part-Aware Spatial–Temporal Graph Convolutional Network, which captures spatial–temporal dependencies among human body key joints and the Predicted Score Inhibition, which refines the output by mitigating the effects of noisy data. By integrating these components, the model enhances robustness in occluded scenarios. The experiments demonstrate that the proposed method outperforms state-of-the-art models on several benchmark datasets, achieving a 6% improvement in recognition accuracy compared to previous approaches. Additionally, we extracted multi-modal features to construct more discriminative features, such as key-joint coordinates, relative coordinates, and temporal differences.