改进LSTM框架下基于动作识别的技战术分析模型  

The technical and tactical analysis model design based on action recognition under the improved LSTM framework

作　　者：李宁 任文[1] 赵智军[1] LI Ning;REN Wen;ZHAO Zhi-jun(Shaanxi Railway Institute,Weinan 714000,Shaanxi Province,China)

机构地区：[1]陕西铁路工程职业技术学院,陕西渭南714000

出　　处：《信息技术》2025年第4期83-88,共6页Information Technology

基　　金：中共陕西省委教育工作委员会研究课题(2021XXM30)。

摘　　要：为了能更好地识别并分析运动员的动作和技战术,文中提出了一种基于3D CNN的多层双向LSTM动作识别分析模型。该模型通过三维卷积神经网络来提取运动特征,采用双向多层LSTM代替传统的单层LSTM来处理时序信息,同时还引入了空间-时间注意力机制,并使用SGD优化器更新模型的权重以减小损失函数。对比同类型的单层LSTM模型和SVM模型,所提模型的F1分数指标分别提升了17%和6.1%,mAP精确度指标提升了2.7%和6.6%,MAE误差降低了49%和44%。经验证,该模型具有较高的可靠性,能够用于运动技战术分析。In order to better identify and analyze athletes’movements and techniques,this paper proposes a multi-layer bidirectional LSTM motion recognition and analysis model based on 3D CNN.This model extracts motion features through a three-dimensional convolutional neural network,uses bidirectional multi-layer LSTM instead of traditional single-layer LSTM to process temporal information,and also introduces a spatio temporal attention mechanism.The SGD optimizer is used to update the model’s weights to reduce the loss function.Compared with the same type of single-layer LSTM model and SVM model,the F1 score index of the proposed model increase 17%and 6.1%,the mAP accuracy index increase 2.7%and 6.6%,and the MAE error decrease 49%and 44%,respectively.After verification,the model has high reliability and can be used for sports technical and tactical analysis.

关 键 词：动作识别 3D CNN 双向LSTM 空间-时间注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]