基于视觉自注意力模型与轨迹滤波器的篮球战术识别

Recognition of Basketball Tactics Based on Vision Transformer and Track Filter

通过机器学习分析球员轨迹数据获得进攻或防守战术,是篮球视频内容理解的关键组成部分。传统机器学习方法需要人为设定特征变量,灵活性大大降低,因此如何自动获取可用于战术识别的特征信息成为关键问题。为此,该文基于美国职业篮球联赛(NBA)比赛中球员轨迹数据设计了一个篮球战术识别模型(TacViT),该模型以视觉自注意力模型(ViT)作为主干网络,利用多头注意力模块提取丰富的全局轨迹特征信息,同时并入轨迹滤波器来加强球场线与球员轨迹之间的特征信息交互,增强球员位置特征表示,其中轨迹滤波器以对数线性复杂度学习频域中的长期空间相关性。该文将运动视觉系统(SportVU)的序列数据转化为轨迹图,自建篮球战术数据集(PlayersTrack),在该数据集上的实验表明,TacViT的准确率达到了82.5%,相对未做更改的视觉自注意力S模型(ViT-S),精度上提升了16.7%。

The analysis of player trajectory data using machine learning to obtain offensive or defensive tactics is a crucial component of understanding basketball video content.Traditional machine learning methods require the setting of feature variables manually,significantly reducing flexibility.Therefore,the key issue is how to automatically obtain feature information that can be used for tactic recognition.To address this issue,a basketball Tactic Vision Transformer(TacViT)recognition model is proposed based on player trajectory data from the National Basketball Association(NBA)games.The proposed model adopts Vision Transformer(ViT)as the backbone network and multi-head attention modules to extract rich global trajectory feature information.Trajectory filters are also incorporated in order to not only enhance the feature interaction between the court lines and player trajectories,but also strengthen the representation of player position features in this study.The trajectory filters learn the long-term spatial correlations in the frequency domain with log-linear complexity.A self-built basketball tactic dataset(PlayersTrack)is created from the sequence data of the Sport Vision System(SportVU),which are converted into trajectory graphs in this work.The experiments on this dataset showed that the accuracy of TacViT reached 82.5%,which is a 16.7%improvement over the accuracy of the Vision Transformer S model(ViT-S)without modifications.