区域时间变化学习的行为识别  

Regional temporal changes learning for action recognition

作　　者：杨兴明[1] 徐浩 汪智文 高旭杰 吴克伟[1] 谢昭[1] Yang Xingming;Xu Hao;Wang Zhiwen;Gao Xujie;Wu Kewei;Xie Zhao(School of Computer Science&Information Engineering,Hefei University of Technology,Hefei 230601,China)

机构地区：[1]合肥工业大学计算机与信息学院,合肥230601

出　　处：《计算机应用研究》2024年第11期3495-3501,共7页Application Research of Computers

基　　金：安徽省自然科学基金资助项目(2108085MF203)。

摘　　要：针对现有的行为识别方法缺少对视频帧中区域级特征的学习,造成识别过程中对相似的行为类别混淆的问题,提出一种区域级时间变化网络。该网络包括局部-全局时间特征学习模块、区域语义学习模块、区域语义融合模块。局部-全局时间特征学习模块学习局部时间注意力,以增强局部视频帧的运动特征,并将其聚合为全局时间区域特征。区域语义学习模块通过计算区域中像素之间的相似度来构建可变化的区域语义卷积核,从而学习随时间变化的行为语义特征。区域语义融合模块将可变化区域特征和全局时间区域特征作为两个独立分支,分别学习每个分支特征的通道注意力用于特征融合。在Something-Something V1&V2与Kinetics-400数据集上的实验结果显示,区域级时间变化网络表现优于多数行为识别方法,证明了该网络能够有效提升行为识别的性能。To solve the problem that existing action recognition methods lack the learning of regional-aware features in video frames,resulting in the confusion of similar action categories in the recognition process,this paper proposed a regional-aware temporal change network.This network included a local-global temporal feature learning module,a regional semantic lear-ning module,and a regional semantic fusion module.The local-global temporal feature learning module learned local temporal attention to enhance video frame features and aggregated them into global temporal region features.The regional semantic learning module constructed changeable region semantic convolution kernels by computing the similarity between pixels in the region to learn action semantic features over time.The regional semantic fusion module took the changeable regional features and global temporal regional features as two independent branches and learned the channel attention of each branch separately for feature fusion.Experiments on the Something-Something V1&V2 and Kinetics-400 datasets show that the regional-aware temporal change network performs better than most action recognition methods,proving that the network can effectively improve the performance of action recognition.

关 键 词：行为识别 区域级特征 卷积神经网络 深度学习 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]