| 摘要: |
| 针对当前人体行为识别准确率低、计算量大等缺陷，提出了一种基于三维卷积神经网络（3D Convolutionnal Neural Network，3D-CNN）结合批量归一化（Batch Normalization，BN）及改进的视觉几何组(Visual Geometry Group,VGG)网络的行为识别算法。该算法首先对3D-CNN结构进行优化，在三维空间采用了多个小卷积核卷积层堆叠的Block结构；同时在网络结构中引入BN算法，将卷积层输出的特征图每一维进行独立的批量归一化处理；之后在Block结构中增加了Dropout层以提高网络泛化能力以及将3D-CNN网络层数加深到了13层，提高了高层次抽象特征的提取能力；最终使用softmax进行分类得出结果。实验结果表明所设计的3D-BN-VGG网络在行为识别方面有较高的识别率。 |
| 关键词: 行为识别；深度学习；三维卷积神经网络；批量归一化 视觉几何组 |
| DOI： |
|
| 基金项目:国家自然科学基金资助项目（61772417，61834005，61802304，61634004，61602377）；陕西省重点研发计划（2017GY-060）；陕西省自然科学基础研究计划项目（2018JM4018） |
|
| A Behavior Recognition Algorithm Based on 3D-BN-VGG Network |
| WU Jin,LI Cong,LIANG Shuang,MIN Yu,WU Hanning |
| (School of Electronic Engineering,Xi′an University of Posts and Telecommunications,Xi′an 710121,China) |
| Abstract: |
| For the defects such as low accuracy and large calculation amount for current human behavior recognition,a behavior recognition algorithm based on 3D convolutionnal neural network(3D-CNN) combined with batch normalization(BN) and improved visual geometry group(VGG) network is proposed.The algorithm first optimizes the 3D-CNN structure,adopts multiple Block structures of small convolutional kernel convolution layer stacks in 3D space,and introduces BN algorithm into the network structure,and outputs the characteristic map of the convolutional layer.The dimension is independently normalized in batch,and then the Dropout layer is added to the Block structure to improve the network generalization ability and deepen the 3D-CNN network layer to 13 layers,which improves the extraction ability of high-level abstract features.The result is finally classified using softmax.The experimental results show that the designed 3D-BN-VGG network has a higher recognition rate in behavior recognition. |
| Key words: behavior recognition deep learning three-dimensional convolutional neural network(3D-CNN) batch normalization visual geometry group |
