| 摘要: |
| 现有的行为识别算法不能充分地提取抽象的行为特征，为此提出了基于三维残差卷积神经网络（3D Residual Convolutional Neural Network，R3D）的人体行为识别算法。该网络在三维卷积神经网络（3D Convolutional Neural Network，3D-CNN）基础上加入了残差模块，可以更好地提取时空域的特征，然后通过改变步长大小进行特征图降维，提高网络效率，并加入批量归一化层和Softplus激活函数，提高网络的收敛速度和拟合能力；之后添加Dropout层，降低过拟合风险，并且使用全局平均池化层(Global Average Pooling，GAP)代替全连接层，克服了网络参数量过大的问题；最后，使用Softmax进行分类。实验结果表明，使用R3D网络在HMDB-51数据集上获得了62.3%的识别率。 |
| 关键词: 行为识别 三维残差卷积神经网络 批量归一化层 全局平均池化层 |
| DOI： |
|
| 基金项目:国家自然科学基金资助项目（61834005，61772417，61602377，61634004）；陕西省重点研发计划项目（2017GY-060）；陕西省自然科学基础研究计划项目（2018JM4018） |
|
| A Human Behavior Recognition Algorithm Based on R3D Network |
| WU Jin,AN Yiyuan,DAI Wei |
| (School of Electronic Engineering,Xi′an University of Posts and Telecommunications,Xi′an 710121,China) |
| Abstract: |
| In view of the problem that the existing behavior recognition algorithms can not extract abstract behavior features fully,a human behavior recognition algorithm based on 3D residual convolutional neural network(R3D) is proposed.Because residual module is added,this network based on 3D convolutional neural network(3D-CNN) can better extract the features of space-time domain,then reduces the dimension of feature map by changing the step size and improves the efficiency of the network.And batch normalization layer and Softplus activation function are added to improve the convergence speed and fitting ability of the network.Then Dropout layer is added to reduce the risk of over fitting,and global average pooling(GAP) instead of full connection layer is used to overcome the problem of too large network parameters.Finally,Softmax is used for classification.The experimental results show that the recognition rate of HMDB-51 dataset is 62.3% by using R3D network. |
| Key words: behavior recognition 3D residual convolutional neural network batch normalization layer global average pooling layer |
《电讯技术》编辑部 