摘要
针对已有的动作识别方法的特征提取不足、识别率较低等问题,结合双流网络、3D卷积神经网络和卷积LSTM网络的优势,提出一种融合模型.该融合模型为了更好地提取人体动作特征,采用SSD目标检测方法将人体目标分割出作为局部特征和原视频的全局特征共同训练,并采用后期融合进行分类;将3D卷积块注意模块采用shortcut结构的方式融合到3D卷积神经网络中,加强神经网络对视频的通道和空间特征提取;并且通过将神经网络中部分3D卷积层替换为ConvLSTM层的方法,更好地得到视频的时序关系.实验在公开的KTH数据集上进行.结果表明,所提模型具有较高的人体动作识别准确率.
Aiming at the problems of insufficient feature extraction and low recognition rate of existing action recognition methods,the paper proposes a fusion model by combining the advantages of two-stream network,3D convolutional neural network and convolutional LSTM network.In order to better extract human motion features,the fusion model adopts SSD target detection method to segment the human body as local features and global features of the original video for joint training,and adopts late fusion for classification.The 3D convolutional block attention module(3D-CBAM)is integrated into 3D convolutional neural network by using shortcut structure to enhance the neural network’s channel and spatial feature extraction.And by replacing part of the 3D convolutional layer of the neural network with ConvLSTM layer,the temporal relation of the video is better obtained.The experiment is carried out on the KTH dataset,and the results show that the proposed model has high recognition accuracy of human action.
作者
王飞
胡荣林
金鹰
Wang Fei;Hu Ronglin;Jin Ying(School of Computer and Software Engineering,Huaiyin Institute of Technology,Huaian 223003,China)
出处
《南京师范大学学报（工程技术版）》
CAS
2021年第1期49-56,共8页
Journal of Nanjing Normal University(Engineering and Technology Edition)
关键词
机器视觉
人体动作识别
3D卷积神经网络
注意力机制
machine vision
human movement recognition
3D convolutional neural network
attention mechanism
作者简介
通讯作者:胡荣林,博士,副教授,研究方向:人机交互技术.E-mail:huronglin@hyit.edu.cn。
