Article: Real-time lightweight pose estimation for sports motion analysis on mobile and edge platforms Journal: International Journal of Business Intelligence and Data Mining (IJBIDM) 2026 Vol.28 No.4/5/6 pp.400 - 416 Abstract: This paper proposes a lightweight motion posture recognition method tailored for real-time sports motion analysis on mobile and edge platforms, which adopts MobileNetV2 as the backbone, augmented with depth-wise separable convolutions and depth-wise transpose convolutions to ensure high-resolution feature up-sampling. A skip concatenation mechanism is introduced to preserve fine-grained spatial information, thereby improving the accuracy of athlete keypoint localisation. To enable real-time feedback in sports scenarios, a full-dataflow pipelining strategy is designed to optimise concurrent operations across feature extraction, pose estimation, and post-processing. Loop unrolling and double buffering are applied to maximise parallelism and minimise memory latency. The proposed approach outperforms state-of-the-art lightweight models on the COCO Keypoints and MPII Human Pose benchmarks. Additionally, it delivers real-time processing speeds of 68 FPS on mobile devices and 180 FPS on edge accelerators. With a compact model size of only 5.2 MB, it ensures efficient deployment in resource-constrained environments. Inderscience Publishers - linking academia, business and industry through research

Title: Real-time lightweight pose estimation for sports motion analysis on mobile and edge platforms

Authors: Haibo Wang; Ningning Li; Chao Liu; Bin Wu

Addresses: Department of Physical Education, Wuxi University of Technology, Wuxi, 214121, Jiangsu, China ' School of Media and Art Design, Wuxi City College of Vocational Technology, Wuxi, 214153, Jiangsu, China ' Department of Physical Education, Wuxi University of Technology, Wuxi, 214121, Jiangsu, China ' School of Integrated Circuits, Wuxi University of Technology, Wuxi, 214121, Jiangsu, China

Abstract: This paper proposes a lightweight motion posture recognition method tailored for real-time sports motion analysis on mobile and edge platforms, which adopts MobileNetV2 as the backbone, augmented with depth-wise separable convolutions and depth-wise transpose convolutions to ensure high-resolution feature up-sampling. A skip concatenation mechanism is introduced to preserve fine-grained spatial information, thereby improving the accuracy of athlete keypoint localisation. To enable real-time feedback in sports scenarios, a full-dataflow pipelining strategy is designed to optimise concurrent operations across feature extraction, pose estimation, and post-processing. Loop unrolling and double buffering are applied to maximise parallelism and minimise memory latency. The proposed approach outperforms state-of-the-art lightweight models on the COCO Keypoints and MPII Human Pose benchmarks. Additionally, it delivers real-time processing speeds of 68 FPS on mobile devices and 180 FPS on edge accelerators. With a compact model size of only 5.2 MB, it ensures efficient deployment in resource-constrained environments.

Keywords: lightweight pose estimation; mobile motion recognition; real-time processing; depth-wise separable convolutions; DwTConv.

DOI: 10.1504/IJBIDM.2026.155253

International Journal of Business Intelligence and Data Mining, 2026 Vol.28 No.4/5/6, pp.400 - 416

Received: 17 May 2025
Accepted: 30 Sep 2025
Published online: 29 Jul 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Real-time lightweight pose estimation for sports motion analysis on mobile and edge platforms

Keep up-to-date