Article: Design of lightweight human pose estimation network for rehabilitation training Journal: International Journal of Business Intelligence and Data Mining (IJBIDM) 2026 Vol.28 No.10 pp.85 - 102 Abstract: Given the resource constraints of terminal devices, existing pose estimation networks, characterised by high computational complexity, struggle to satisfy the simultaneous demands of real-time performance and accuracy in rehabilitation training. Therefore, this paper proposes a lightweight human pose estimation network specifically designed for rehabilitation training. Firstly, laser triangulation is used to obtain the target image, and precise mapping from 3D to 2D space is achieved through coordinate transformation. Secondly, design an RMPE tiny lightweight network, introduce G-Bottleneck module to compress parameter quantity, and integrate Sa-ECA attention mechanism to enhance feature interaction. Finally, the Huber loss function is used to optimise training and improve convergence speed and robustness. The experimental results show that the method proposed in this paper maintains an overall accuracy of over 96% in human pose estimation testing, with most samples approaching or exceeding 98%, consistently maintains between 41.1 FPS and 44.4 FPS in ten inference speed tests. Inderscience Publishers - linking academia, business and industry through research

Title: Design of lightweight human pose estimation network for rehabilitation training

Authors: Qingyun Yang; Gaigai Zhang

Addresses: College of Rehabilitation Medicine, North Henan Medical University, He nan, 453003, China ' Department of Intelligent Medical Engineering, North Henan Medical University, He nan, 453003, China

Abstract: Given the resource constraints of terminal devices, existing pose estimation networks, characterised by high computational complexity, struggle to satisfy the simultaneous demands of real-time performance and accuracy in rehabilitation training. Therefore, this paper proposes a lightweight human pose estimation network specifically designed for rehabilitation training. Firstly, laser triangulation is used to obtain the target image, and precise mapping from 3D to 2D space is achieved through coordinate transformation. Secondly, design an RMPE tiny lightweight network, introduce G-Bottleneck module to compress parameter quantity, and integrate Sa-ECA attention mechanism to enhance feature interaction. Finally, the Huber loss function is used to optimise training and improve convergence speed and robustness. The experimental results show that the method proposed in this paper maintains an overall accuracy of over 96% in human pose estimation testing, with most samples approaching or exceeding 98%, consistently maintains between 41.1 FPS and 44.4 FPS in ten inference speed tests.

Keywords: rehabilitation training; lightweight network; human pose estimation; attention mechanism.

DOI: 10.1504/IJBIDM.2026.156184

International Journal of Business Intelligence and Data Mining, 2026 Vol.28 No.10, pp.85 - 102

Received: 23 Mar 2026
Accepted: 02 Jun 2026
Published online: 07 Sep 2026 *

Title: Design of lightweight human pose estimation network for rehabilitation training

Keep up-to-date