A Badminton Stroke Type Recognition Model based on Machine Learning
收藏资源简介：
The article Aims to compare the application effect of different machine learning algorithms in the recognition of badminton stroke types, so as to better meet the needs of quantitative analysis of badminton skills and tactics, and provide a more reasonable basis for the layout of game strategies. We used wearable inertial sensor device on the wrist to collect a total of 708 shots of linear acceleration and angular velocity data, and extract the domain features including data range, minimum maximum, average, absolute average, kurtosis, skewness statistics and P value, entropy, standard deviation, Angle range, quartile spacing, maximum minimum relative position. Then, we labeled each hit type data manually, total divided into the forehand overarm, backhand overarm, forehand underarm, backhand underarm, forehand smash. The results show that: (1) K-nearest neighbors and k-means clustering show poor prediction accuracy and application accuracy; (2) Logistic regression shows great accuracy in recognizing the overarm stroke, while random forest performs relatively well in recognizing underarm stroke; (3) The angular velocity data of the gyroscope is the characteristic that affects the classification recognition effect, and the difference in stroke type shown in wrist is mainly the angular velocity change.
本研究旨在对比不同机器学习算法在羽毛球击球类型识别中的应用效果，以更好地适配羽毛球技战术定量分析的需求，为赛事战略布局提供更为合理的依据。 本研究通过佩戴于手腕的可穿戴惯性传感器设备（Wearable Inertial Sensor Device），共采集708次击球的线性加速度与角速度数据，并提取了涵盖数据极差、最值、均值、绝对均值、峰度、偏度统计量、P值、熵、标准差、角度极差、四分位距及最值相对位置在内的领域特征。 随后，研究人员对每一条击球类型数据进行人工标注，共划分为正手上手击球、反手上手击球、正手下手击球、反手下手击球及正手扣杀五类。 研究结果表明：（1）K近邻（K-nearest Neighbors）算法与K均值聚类（K-means Clustering）模型的预测精度与应用效果均较差；（2）逻辑回归（Logistic Regression）在识别上手击球类型时表现出优异的准确率，而随机森林（Random Forest）在下手击球类型的识别任务中性能相对更佳；（3）陀螺仪（Gyroscope）采集的角速度数据是影响分类识别效果的核心特征，手腕处展现的击球类型差异主要体现为角速度变化。
