王粉花^{1, 2, 3,,},
张强¹,
黄超¹,
张苒¹
1.北京科技大学自动化学院 北京 100083
2.北京科技大学人工智能研究院 北京 100083
3.北京市工业波谱成像工程中心 北京 100083
基金项目:国家重点研发计划重点专项(2017YFB1400101-01)，北京科技大学中央高校基本科研业务费专项资金(FRF-BD-19-002A)

详细信息

作者简介:王粉花：女，1971年生，博士，副教授，硕士生导师，研究方向为模式识别与智能信息处理
张强：男，1994年生，硕士生，研究方向为图像处理与手势识别
黄超：男，1993年生，硕士生，研究方向为图像处理

通讯作者:王粉花　wangfenhua@ustb.edu.cn

中图分类号:TP183

计量

文章访问数:681
HTML全文浏览量:246
PDF下载量:133
被引次数:0

出版历程

收稿日期:2020-01-16
修回日期:2020-12-06
网络出版日期:2020-12-18
刊出日期:2021-05-18

Dynamic Gesture Recognition Combining Two-stream 3D Convolution with Attention Mechanisms

Fenhua WANG^{1, 2, 3,,},
Qiang ZHANG¹,
Chao HUANG¹,
Ran ZHANG¹
1. School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, China
2. Institute of Artificial Intelligence, University of Science and Technology Beijing, Beijing 100083, China
3. Beijing Engineering Research Center of Industrial Spectrum Imaging, Beijing 100083, China
Funds:The National Key Research and Development Project of China (2017YFB1400101-01), The Fundamental Research Funds for the Central Universities (FRF-BD-19-002A)


摘要
摘要:得益于计算机硬件以及计算能力的进步，自然、简单的动态手势识别在人机交互方面备受关注。针对人机交互中对动态手势识别准确率的要求，该文提出一种融合双流3维卷积神经网络(I3D)和注意力机制(CBAM)的动态手势识别方法CBAM-I3D。并且改进了I3D网络模型的相关参数和结构，为了提高模型的收敛速度和稳定性，使用了批量归一化(BN)技术优化网络，使优化后网络的训练时间缩短。同时与多种双流3D卷积方法在开源中国手语数据集(CSL)上进行了实验对比，实验结果表明，该文所提方法能很好地识别动态手势，识别率达到了90.76%，高于其他动态手势识别方法，验证了所提方法的有效性和可行性。
关键词:动态手势识别/
深度学习/
双流3维卷积神经网络/
注意力机制/
BN层
Abstract:Benefits from the progress of computer hardware and computing power, natural and simple dynamic gesture recognition gets a lot of attention in human-computer interaction. In view of the requirement of the accuracy of dynamic gesture recognition in human-computer interaction, a method of dynamic gesture recognition that combines Two-stream Inflated 3D (I3D) Convolution Neural Network (CNN) with the Convolutional Block Attention Module (CBAM-I3D) is proposed. In addition, relevant parameters and structures of the I3D network model are improved. In order to improve the convergence speed and stability of the model, the Batch Normalization (BN) technology is used to optimize the network, which shortens the training time of the optimized network. At the same time, experimental comparisons with various Two-stream 3D convolution methods on the open source Chinese Sign Language (CSL) recognition dataset are performed. The experimental results show that the proposed method can recognize dynamic gestures well, and the recognition rate reaches 90.76%, which is higher than other dynamic gesture recognition methods. The validity and feasibility of the proposed method are verified.
Key words:Dynamic gesture recognition/
Deep learning/
Two-stream 3D Convolution Neural Network (CNN)/
Attention mechanism/
Batch Normalization (BN) layer



PDF全文下载地址:

https://jeit.ac.cn/article/exportPdf?id=e6c137fe-1cc5-4b87-a078-51c79f925580