星际争霸图像序列的群体行为识别研究  

Research on group behavior recognition of starcraft image sequence

作　　者：白江波 杨阳[1] 张文生[1] BAI Jiangbo;YANG Yang;ZHANG Wensheng(Institute of Automation,Chinese Academy of Sciences,Beijing 100080,China)

机构地区：[1]中国科学院自动化研究所,北京100080

出　　处：《应用科技》2022年第2期63-69,共7页Applied Science and Technology

基　　金：广东省重点领域研发计划项目(2020B0101130020);国家自然科学基金项目(U1936206,61906191)。

摘　　要：以星际争霸游戏图像数据的群体目标行为识别为研究目标,构建游戏中群体目标典型行为的数据集,并应用当前主流的行为识别方法对游戏图像序列进行了群体行为识别。群体目标典型行为的数据集共包含战斗、移动、静止这3类典型行为,每个行为样本中包含不同数量、不同背景和不同行为表示形式的目标单元,充分保证了数据样本的多样性。分别使用时间段网络(TSN)、3D卷积(C3D)网络、膨胀3D卷积网络(I3D)以及时间移位模块(TSM)这4种行为识别方法,对构建的星际争霸图像序列群体目标行为识别数据集进行了验证。实验结果表明,本数据集有效,且基于3D卷积网络的C3D模型取得了最好的行为识别效果和最短的计算时间,满足群体目标行为识别任务需求。本文为后续研究遥感图像的群体目标行为识别方法提供了指导和迁移学习的原始数据。This paper takes the group target behavior recognition of StarCraft game image data as the research target,constructs dataset for the typical behaviors of group targets in the game, and applies the current mainstream behavior recognition methods to recognize the group behavior in game image sequences. The dataset for the typical behaviors of group targets includes three types of typical behaviors: fighting, moving, and stationary. Each behavior sample contains a different number of target units with different backgrounds and behavior representations, which fully guarantees the diversity of data samples. This paper uses four behavior recognition methods(temporal segment networks(TSN), 3D convolution(C3D), inflated 3D convNet(I3D), temporal shift module(TSM)) to verify the dataset. The experimental results show that this dataset is effective, and that the C3D model based on 3D convolution network achieves the best behavior recognition effect and consumes the shortest calculation time, meeting the needs of group target behavior recognition tasks. This article provides guidance and original data for transfer learning for subsequent research on the method of group target behavior recognition in remote sensing images.

关 键 词：图像序列 群体行为识别 3D卷积 星际争霸 行为数据集 神经网络 深度学习 残差网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]