基于机器学习的羽毛球挥拍动作识别算法比较研究

Comparative study of badminton swing motion recognition algorithms based on machine learning

尽管人们对羽毛球运动关注度较高,但是面向羽毛球运动的智能化设备却很少。本文基于传统的浅层机器学习随机森林(Random Forest,RF)、K近邻(K Nearest Neighbor,K NN)、梯度提升(Gradient Boosting,GB)、支持向量机(Support Vector Machine,SVM)和长短期记忆网络(Long Short Term Memory,LSTM)深度学习模型验证了能够准确识别头顶正手击球、架空反手击球、杀球、腋下正手击球、腋下反手击球5种常见的羽毛球挥拍动作类型的算法。本研究通过固定在球拍握把底部的无线惯性传感器模块采集了12名运动者的挥拍动作数据样本,共1800组,采用低功耗蓝牙进行数据传输和收集,实验过程中采集的数据通过动作窗口和滑动窗口相结合的窗口切割方法进行截取,提取经窗口截取后的动作数据特征,使用RF、K NN、GB、SVM和LSTM模型学习验证识别了实验中的5种挥拍动作。实验结果表明,LSTM识别精度达到99.42%,明显优于传统的机器学习算法。同时,本文选择STM32F476 ARM微控制器作为边缘计算单元,将基于LSTM的羽毛球挥拍动作识别模型部署到该微控制器中,用于实时推断和识别运动者羽毛球挥拍动作类型,识别效果良好。

Despite the intense attention to badminton,there are relatively few intelligent devices specifically designed for this sport.Therefore,this study verifies algorithms based on traditional shallow machine learning models such as Random Forest(RF),K-Nearest Neighbor(K NN),Gradient Boosting(GB),Support Vector Machine(SVM),and Long Short-Term Memory(LSTM)deep learning models to accurately identify five common badminton swing actions:overhead forehand stroke,aerial backhand stroke,smash,underarm forehand strokes,and underarm backhand stroke.This study collected 1800 sets of swing motion data samples from 12 athletes using a wireless inertial sensor module fixed at the bottom of the badminton racket grip.Low power Bluetooth was used for data transmission and collection.The real-time data collected was intercepted using a window cutting method that combines action window and sliding window.The feature of the intercepted action data was extracted.RF,K NN,GB,SVM,and LSTM models were used to learn and verify the recognition of five swing movements during the experiment.The experimental results showed that LSTM reached a recognition accuracy of 99.42%,significantly outperforming traditional machine learning algorithms.Additionally,this paper selects STM32F476 ARM microcontroller as the edge computing unit,and deploys the badminton swing action recognition model into it.This deployment enables real-time inference and recognition of badminton swing types by athletes,demonstrating effective recognition performance.