摘要
在多人在线战术竞技游戏(MOBA)中,分析对手的行为模式能够极大提高电子竞技胜率。行为识别是行为模式分析的前提,因此准确并高效地识别电子竞技行为成为了一个急需解决的问题。目前在电子竞技中解决行为识别问题主要使用人工标注的方法,但是基于人工标注的方法效率极低。而使用深度学习解决行为识别问题又因为样本数据量过小导致准确率较低。针对以上问题,本研究提出一种基于稠密关系蒸馏和元特征选择的小样本识别网络,采用上下文感知的特征聚合机制识别关键的行为,同时引入多尺度Harris图像配准算法,利用电子竞技地图强对称的特点,提高了有效数据量。实验结果表明,本研究提出的方法有着较高的识别准确率与效率。
In Multiplayer Online Battle Arena Games(MOBA),analyzing the rivals’behavior can greatly improve the winning rate of e-sports.The premise of behavior pattern analysis is behavior recognition,so accurate and efficient identification of e-sports behavior has become a problem to be solved.At present,in e-sports,the method of manual annotation was mainly used to solve the problem of behavior recognition,but the efficiency of the method based on manual annotation was very low.The use of deep learning to solve behavior recognition problems resulted in low accuracy due to the small sample data size.To solve the above problems,a small sample recognition network based on dense relation distillation and meta feature selection was proposed in this study,which uses context aware feature aggregation mechanism to identify key behaviors.At the same time,the multi-scale Harris image registration algorithm was introduced to improve the effective data volume by taking advantage of the symmetrical characteristics of the electronic sports map.The experimental results showed that the method proposed in this study has high recognition accuracy and efficiency.
作者
王欣
徐平平
徐蕾艳
WANG Xin;XU Ping-ping;XU Lei-yan(School of Digital Business,Nanjing Vocational College of Information Technology,Nanjing 210023,China;School of Information Science and Engineering,Southeast University,Nanjing 210096,China)
出处
《印刷与数字媒体技术研究》
CAS
北大核心
2023年第6期99-106,共8页
Printing and Digital Media Technology Study
基金
国家自然科学基金(No.12101319)
江苏省产学研合作项目(No.BY2022940)
南京信息职业技术学院高层次人才科研启动项目(No.YB20220602)。
