基于注意力机制的改进残差网络的人体行为识别方法
2021-11-09王昊飞李俊峰
王昊飞 李俊峰



摘 要:针对ResNeXt网络(残差网络)中存在的对特征提取不充分,以及数据集中背景信息干扰的问题,将ResNeXt网络和注意力机制相结合,提出了一种基于注意力机制的ResNeXt模型。首先,在ResNeXt网络的基础上,将浅层和深层的特征融合生成新型网络结构。其次,将全连接层由全局平均池化层替代,然后在通道空间注意力机制中添加一个条件因子,同时将改进后的注意力机制嵌入上述网络中。最后,在UCF101和HMDB51上分别进行实验,得到了95.2%和65.6%的准确率。研究表明,本文提出的模型可以有效地提取关键特征,充分利用不同层次的特征信息获得较好的准确率。
关键词:人体行为识别;注意力机制;ResNeXt;全局平均池化
中图分类号:TP183 文献标识码:A
Human Action Recognition Method based on Attention
Mechanism and Improved ResNeXt Network
WANG Haofei, LI Junfeng
(Faculty of Mechanical Engineering & Automation, Zhejiang Sci-Tech University, Hangzhou 310018, China)
haofeiwang@yeah.net; ljf2003zz@163.com
Abstract: Aiming at problems of insufficient feature extraction in ResNeXt network and background information interference in the dataset, this paper proposes a ResNeXt model based on attention mechanism, which combines the ResNeXt network and attention mechanism. First, based on ResNeXt network, shallow and deep features are merged to generate a new network structure. Second, the fully connected layer is replaced by a global average pooling layer. Then channel attention mechanism is improved by adding a condition factor. At the same time, the improved attention mechanism is embedded in the above-mentioned network. Finally, experiments are performed on UCF101 and HMDB51 respectively, and the accuracy rates of 95.2% and 65.6% are obtained. Experiments show that the proposed model can effectively extract key features, and make full use of feature information of different layers to achieve better accuracy.
Keywords: human action recognition; attention mechanism; ResNeXt network; global average pooling
1 引 言(Introduction)
人體行为识别技术是从包含运动信息的图像、视频中进行识别的。在视频监控、智能家居、运动分析以及VR等领域都离不开人体行为的识别。人体行为识别已成为计算机视觉研究中的一个非常重要的领域[1]。由于视点的不同、背景的复杂性以及光照条件等的影响,人体行为识别仍然是一项非常具有挑战性的课题。传统人体行为识别是基于手工设计的特征[2]进行识别,并且依赖数据集特征提取的先验知识,耗费大量的时间和精力。随着深度学习的兴起,解决了手动设计特征的不足,在人体行为识别领域取得了重大进展[3],已经明显超过了手工设计的特征。XIE等[4]提出了ResNeXt网络,用一种平行堆叠相同拓扑结构的blocks来代替残差网络三层卷积的block,同时增加了“基数”这一概念,减少了超参数数量,计算效率高,准确率高。……
