基于堆叠式长短期记忆网络的篮球运动员微动作评价方法  

Evaluation Method of Basketball Player Micro Motions Based on Stackable Long and Short Term Memory Network

作　　者：牛程程 鲁大营 郑亚淼[3] NIU Chengcheng;LU Daying;ZHENG Yamiao(School of General Education,Qingdao Huanghai University,Qingdao 266427,China;School of Software,Qufu Normal University,Jinan 273165,China;School of Physical Education Science,Qufu Normal University,Jinan 273165,China)

机构地区：[1]青岛黄海学院通识教育学院,山东青岛266427 [2]曲阜师范大学软件学院,山东济南273165 [3]曲阜师范大学体育科学学院,山东济南273165

出　　处：《湖南科技大学学报（自然科学版）》2022年第2期95-103,共9页Journal of Hunan University of Science And Technology：Natural Science Edition

基　　金：山东省高等学校科技计划项目资助(J15LN83);山东省社会科学规划研究项目资助(14CTYJ21)。

摘　　要：为了评估运动员微动作在持球回合中的贡献,提出一个端到端深度学习的评价方法.所提方法不需要复杂的特征提取(如状态间转换的定义和建模),而是将球员和篮球的原始轨迹作为输入,利用堆叠式长短期记忆(Long Short-Term Memory,LSTM)网络学习时空窗口的特征表示,通过一个额外的全连接层对场上球员的隐性空间表征进行级联处理.利用Softmax层对球员的终结动作(如投篮得分、失误、犯规等)的概率进行估计,每个终结动作均与一个预期分值关联,并用其估计预期得分.为了解决数据的不平衡性,对训练阶段使用参数化的下采样方案.实验结果表明:所提方法可以准确地估计回合结果的概率分布,对技术统计数据之外的微动作评价具有参考价值.To evaluate the contribution of basketball micro motions in holding round,an evaluation method of end-to-end deep learning was proposed.The proposed method needn’t complex feature extraction(such as the definition and modeling of transition between states),but took the original trajectory of players and basketball as input.The long short term memory(LSTM)network was used to learn the feature representation of spatiotemporal window,and the implicit spatial representation of players on the court was cascaded through an extra full connection layer.The Softmax layer was used to estimate the probability of a player's final actions(such as shooting,mistakes,fouls,etc.),and each finishing action was associated with an expected score,and the expected score was estimated by using it.In order to solve the imbalance of data,a parameterized down sampling scheme was used in the training phase.The experimental results show that the proposed method accurately estimate the probability distribution of turn results,and has reference value for micro motion evaluation beyond technical statistics.

关 键 词：深度学习 长短期记忆网络 时空跟踪数据 篮球微动作 下采样 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]