Scholarlink

This chapter introduces an AI-powered system for real-time recognition of physiotherapy-related exercises to assist with rehabilitation and healthcare. The system comprises two core components: (i) a lightweight deep learning model for human action recognition (HAR) using 3D skeleton keypoints with MediaPipe framework (Lugaresi et al. MediaPipe: a framework for building perception pipelines (2019). arXiv Preprint. arXiv:1906.08172) and (ii) a mobile app enabling real-time interaction between therapists and patients. We compare two neural architectures: Graph Convolutional Network (GCN) and a pose-based transformer (SPOTER), pretrained on the Yoga Pose Dataset and fine-tuned on a custom dataset of 13 participants performing four core exercises. While both models achieved perfect validation accuracy (F1 \(=\) 1.00), GCN is substantially more lightweight (51K parameters) than SPOTER (1.3M). The real-time functionality was demonstrated via a cross-platform mobile prototype. These findings confirm GCN’s potential for embedded HAR applications in clinical settings.

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

Personalized Rehabilitation and Athletic Performance Enhancement Using AI Techniques

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Khanh Duong Bao,
  • Hoang Le Minh,
  • Thao Phan Le Thanh,
  • Nhan Ha The,
  • Nguyen Huu Duc Minh,
  • Trung Nguyen Quoc,
  • Nguyen Van Bay,
  • Duong Huu Thanh

摘要

This chapter introduces an AI-powered system for real-time recognition of physiotherapy-related exercises to assist with rehabilitation and healthcare. The system comprises two core components: (i) a lightweight deep learning model for human action recognition (HAR) using 3D skeleton keypoints with MediaPipe framework (Lugaresi et al. MediaPipe: a framework for building perception pipelines (2019). arXiv Preprint. arXiv:1906.08172) and (ii) a mobile app enabling real-time interaction between therapists and patients. We compare two neural architectures: Graph Convolutional Network (GCN) and a pose-based transformer (SPOTER), pretrained on the Yoga Pose Dataset and fine-tuned on a custom dataset of 13 participants performing four core exercises. While both models achieved perfect validation accuracy (F1 \(=\) 1.00), GCN is substantially more lightweight (51K parameters) than SPOTER (1.3M). The real-time functionality was demonstrated via a cross-platform mobile prototype. These findings confirm GCN’s potential for embedded HAR applications in clinical settings.