融入注意力机制的时间激励与聚集行为识别模型  

Temporal excitation and aggregation model integrated attention mechanism for action recognition

作　　者：李文静 高向军 刘梦雪 葛方振[1] 沈龙凤[1] 刘怀愚[1] LI Wen-jing;GAO Xiang-jun;LIU Meng-xue;GE Fang-zhen;SHEN Long-feng;LIU Huai-yu(School of Computer Science and Technology,Huaibei Normal University,Huaibei 235000,China)

机构地区：[1]淮北师范大学计算机科学与技术学院,安徽淮北235000

出　　处：《云南民族大学学报（自然科学版）》2022年第6期741-748,共8页Journal of Yunnan Minzu University:Natural Sciences Edition

基　　金：安徽省重点研究与开发计划面上攻关项目(201904a05020072);安徽省高校自然科学研究项目(KJ2019A0606,KJ2019A0603).

摘　　要：视频行为识别通常应用短区间动作特征和长区间视频聚集特征进行时序建模.而这种时序建模方式在特征提取过程中,将不同时序区间的动作特征同等看待,忽略了关键通道信息和重要动作内容,不能达到理想的行为识别效果.注意力机制能够重点关注目标区间,在提取时间激励与聚集行为特征基础上融入通道-空间注意力模块.该模型分别通过通道和空间模块改变时序动作的特征分布,通道注意力关注关键通道信息是“哪些”,空间注意力机制关注重要视频内容在“哪里”,突出关键通道信息和重要内容等特征,提高了行为识别的识别率.同时在数据集Something-Something 1,UCF101和HMDB51对模型进行实验,融入通道-空间注意力模块的时间激励与聚集行为识别模型能够有效提高行为识别率.Video action recognition usually uses short-range motion features and long-range video gathering features to conduct temporal modeling.However,in the process of feature extraction,this temporal modeling method treats the action features of different temporal interval equally,ignores the key channel information and important action content,and fails to achieve ideal action recognition effect.Attention mechanism is able to focus on target range.On the basis of extracting temporal excitation and aggregation features,the channel-spatial attention module was devised.The model changes the timing characteristics of distribution through the movement of the channel and spatial module.The channel attention key information is“what,and spatial attention mechanism focuses on where”the the important video content and key channel information are.Important content is highlighted to improve the recognition rate of action recognition.In this paper,the model is tested on the data sets something-SomethingV1,UCF101 and HMDB51,and the temporal excitation and aggregation action recognition model integrated with the channel-spatial attention module can effectively improve the action recognition rate.

关 键 词：行为识别 卷积神经网络 时间激励与聚集 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]