Vision-based human pose estimation for intelligent sports training and teaching assistance
TL;DR
GSANet substantially reduces model parameters and computational complexity relative to the reported HRNet-W32 configuration without ImageNet pre-training, although its absolute pose-estimation accuracy is lower.
Abstract
Vision-based human pose estimation can support intelligent sports training and teaching, but fast motion, motion blur, occlusion, background interference, and computational constraints remain challenging in practical scenarios. This study proposes a sports-oriented framework integrating motion deblurring, target tracking, and a lightweight high-resolution pose estimator termed GSANet. GSANet redesigns HRNet using Ghost modules, Sandglass structures, Coordinate Attention, and unbiased data processing to reduce computational complexity while retaining high-resolution spatial representation. COCO and MPII are used as the principal benchmarks, while a small supplementary sports-oriented set, SGDN, is used to examine the influence of motion blur and preprocessing-based deblurring. GSANet substantially reduces model parameters and computational complexity relative to the reported HRNet-W32 configuration without ImageNet pre-training, although its absolute pose-estimation accuracy is lower. On SGDN, preprocessing-based deblurring produces a modest but statistically supported improvement in PCKh. Overall, the results support GSANet as an efficiency-oriented alternative under the evaluated settings rather than as a universally more accurate or deployment-ready pose estimator.