基于注意力模型的篮球视频事件和关键角色检测方法

DETECTION OF BASKETBALL VIDEO EVENTS AND KEY ROLES BASED ON ATTENTION MODEL

为了对篮球比赛视频中的关键角色和重要事件进行检测,考虑到"注意力"与正在进行的篮球活动高度相关,提出一种基于注意力模型的方法。构建篮球比赛数据集,对11个关键的事件类型进行手工识别;对视频中的运动员进行跟踪,跟踪特征采用双向长短期记忆(Bi-directional Long Short-Term Memory,BLSTM)网络表示,使用注意力模型从输入到输出对元素进行对齐;使用另一个BLSTM对受关注的特征进行处理,以进行事件的检测和分类。实验结果表明,所提方法在事件分类和检测上的性能均优于一些同类方法。另外,除对篮球事件进行识别之外,还能够识别参与事件中的关键球员。

To detect key roles and important events in basketball game video,a method based on attention model is proposed,considering that attention is highly correlated with ongoing basketball activities.The data set of basketball matches was constructed and 11 key event types were identified manually.Players in the video were tracked,and the tracking features were represented by bi-directional long short-term memory(BLSTM)network.The elements were aligned from input to output using the attention model.Another BLSTM was used to process the features of interest for event detection and classification.The experimental results show that the performance of proposed method is better than some similar methods in event classification and detection.In addition,besides identifying basketball events,it can also identify key player involved in events.