Scholarlink

<p>Rehabilitation behavior recognition is critical for monitoring patient progress and providing feedback during physical therapy. This paper introduces a novel Spatio-Temporal Attention-Augmented Gated Graph Convolutional Network (STA-GGCN) designed to capture the complex dynamics of human motion using skeletal data. The STA-GGCN is specifically designed to capture the complex dynamics of human motion using skeletal data. The proposed model integrates graph-based spatial modeling and gated mechanisms to selectively manage joint interactions. Additionally, a spatio-temporal attention module dynamically identifies critical joints and frames, while recurrent layers model temporal dependencies. We evaluated the STA-GGCN on three benchmark datasets: UPFall, Human3.6M, and Kinetics-Skeleton. The model achieved accuracy scores of 89.14%, 90.82%, and 88.31%, respectively, along with high temporal alignment as measured by Temporal Intersection over Union (tIoU). Comprehensive analyses demonstrate the contributions of individual components, the robustness of the model across datasets, and its interpretability through attention heatmaps and Grad-CAM. Despite its computational overhead and sensitivity to noisy skeletal data, STA-GGCN exhibits strong generalization and interpretability, rendering it suitable for applications in rehabilitation, sports analytics, and human-computer interaction.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel spatio-temporal attention-based graph model for patient rehabilitation behavior recognition

  • 期刊论文

  • 发表日期：

    2025年6月20日

查看全文
  • Hang Cao,
  • Qiang Chen,
  • Song Wang

摘要

Rehabilitation behavior recognition is critical for monitoring patient progress and providing feedback during physical therapy. This paper introduces a novel Spatio-Temporal Attention-Augmented Gated Graph Convolutional Network (STA-GGCN) designed to capture the complex dynamics of human motion using skeletal data. The STA-GGCN is specifically designed to capture the complex dynamics of human motion using skeletal data. The proposed model integrates graph-based spatial modeling and gated mechanisms to selectively manage joint interactions. Additionally, a spatio-temporal attention module dynamically identifies critical joints and frames, while recurrent layers model temporal dependencies. We evaluated the STA-GGCN on three benchmark datasets: UPFall, Human3.6M, and Kinetics-Skeleton. The model achieved accuracy scores of 89.14%, 90.82%, and 88.31%, respectively, along with high temporal alignment as measured by Temporal Intersection over Union (tIoU). Comprehensive analyses demonstrate the contributions of individual components, the robustness of the model across datasets, and its interpretability through attention heatmaps and Grad-CAM. Despite its computational overhead and sensitivity to noisy skeletal data, STA-GGCN exhibits strong generalization and interpretability, rendering it suitable for applications in rehabilitation, sports analytics, and human-computer interaction.