Scholarlink

Introduction <p>The field of sports analytics is being transformed by artificial intelligence (AI) and machine learning, which enable objective assessments based on developed modelling methods. The paper presents a k-Nearest Neighbours-Hidden Markov Model (KNN-HMM) to address issues of accuracy, time-dependent characteristics, and flexibility in the study of badminton strokes.</p> Methods <p>The dataset comprised 10,000 motion-captured badminton strokes collected in Chongqing, China, from 36 players (18 male, 18 female; ages 18–35) between 2015 and 2017. Vicon optical motion capture was used to capture the data, which were then annotated in Kinovea. Angular velocity and jerk were also used to extract spatial and temporal features in 200-ms sliding windows containing 50-ms of overlap. The model combined a KNN stroke classifier with HMM-based temporal sequence prediction. Accuracy, precision, recall, and F1-score were the evaluation metrics in the fivefold cross-validation, and real-time metrics were used: latency and throughput.</p> Results <p>The KNN-HMM achieved 97.5% accuracy with a latency of 32&#xa0;ms and 90.8% streaming accuracy. In comparison to standalone KNN, modest performance gains were observed in accuracy (+ 0.05, p = 0.021), F1-score (+ 0.07, p = 0.015), precision (+ 0.06, p = 0.001), and recall (+ 0.04, p = 0.018), with Tukey’s Honestly Significant Difference (HSD) confidence intervals confirming significance. Improvements were evident across several strokes, including Clear, Jump Smash, and Net Shot (p &lt; .05). Compared with alternative models such as hybrid SVM-Decision Tree-KNN (83.4%) and acoustic sensor-based approaches (~ 84%), the KNN-HMM achieved stronger results. The difference of 1.03% compared with personalised acoustic models (96.5%) was small, but it showed strength on a more generalised dataset.</p> Conclusion <p>KNN-HMM can offer an effective paradigm of real-time classification of badminton stroke, but the results are limited to a regional model (Chongqing, China). Although it is based on confined-area data, the model has high accuracy, adaptability, and generalizability, and is advantageous for training applications and real-time tactical analysis.</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

Real-time stroke prediction in badminton integrating AI with the KNN-HMM model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月6日

查看全文
  • Qi Song,
  • Jian Zhang

摘要

Introduction

The field of sports analytics is being transformed by artificial intelligence (AI) and machine learning, which enable objective assessments based on developed modelling methods. The paper presents a k-Nearest Neighbours-Hidden Markov Model (KNN-HMM) to address issues of accuracy, time-dependent characteristics, and flexibility in the study of badminton strokes.

Methods

The dataset comprised 10,000 motion-captured badminton strokes collected in Chongqing, China, from 36 players (18 male, 18 female; ages 18–35) between 2015 and 2017. Vicon optical motion capture was used to capture the data, which were then annotated in Kinovea. Angular velocity and jerk were also used to extract spatial and temporal features in 200-ms sliding windows containing 50-ms of overlap. The model combined a KNN stroke classifier with HMM-based temporal sequence prediction. Accuracy, precision, recall, and F1-score were the evaluation metrics in the fivefold cross-validation, and real-time metrics were used: latency and throughput.

Results

The KNN-HMM achieved 97.5% accuracy with a latency of 32 ms and 90.8% streaming accuracy. In comparison to standalone KNN, modest performance gains were observed in accuracy (+ 0.05, p = 0.021), F1-score (+ 0.07, p = 0.015), precision (+ 0.06, p = 0.001), and recall (+ 0.04, p = 0.018), with Tukey’s Honestly Significant Difference (HSD) confidence intervals confirming significance. Improvements were evident across several strokes, including Clear, Jump Smash, and Net Shot (p < .05). Compared with alternative models such as hybrid SVM-Decision Tree-KNN (83.4%) and acoustic sensor-based approaches (~ 84%), the KNN-HMM achieved stronger results. The difference of 1.03% compared with personalised acoustic models (96.5%) was small, but it showed strength on a more generalised dataset.

Conclusion

KNN-HMM can offer an effective paradigm of real-time classification of badminton stroke, but the results are limited to a regional model (Chongqing, China). Although it is based on confined-area data, the model has high accuracy, adaptability, and generalizability, and is advantageous for training applications and real-time tactical analysis.