计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (2): 100-103.DOI: 10.3778/j.issn.1002-8331.1804-0373

• 模式识别与人工智能 • 上一篇    下一篇

多尺度方法结合卷积神经网络的行为识别

盖  赟1,荆国栋2   

  1. 1.中国社会科学院大学 计算机教研部,北京 100089
    2.中国气象局气象干部培训学院 远程教育中心,北京 100081
  • 出版日期:2019-01-15 发布日期:2019-01-15

Human Action Recognition Based on Convolution Neural Network Combined with Multi-Scale Method

GE Yun1, JING Guodong2   

  1. 1.Department of Computer Teaching and Research, University of Chinese Academy of Social Sciences, Beijing 100089, China
    2.Distance Education Center, China Meteorological Administration Training Centre, Beijing 100081, China
  • Online:2019-01-15 Published:2019-01-15

摘要: 为了同时计算行为序列样本在时间和空间的特征,提出了一种基于包含多尺度卷积算子的卷积神经网络识别模型。首先通过叠加的方式将序列样本中的骨骼向量信息整合为一个行为矩阵,然后将矩阵输入识别模型。为了挖掘具有不同邻接关系的骨骼点在描述人体行为时的作用,将卷积神经网络各层中的卷积算子拓展为多尺度卷积算子,并使用该网络得到的特征进行分类。实验在MSR-Action3D数据集和HDM05数据集获得较好的识别率。

关键词: 行为识别, 时空特征, 深度卷积神经网络, 深度学习, 行为矩阵

Abstract: In order to simultaneously calculate the temporal and spatial feature of behavior sequence samples, a convolution neural network recognition model with multi-scale convolution operator is proposed. Firstly, the skeleton vector in the sequence sample is integrated into a behavior matrix by superposition. Then the matrix is input into the recognition model. The convolution operators in each layers of the convolution neural network are extended to the multi-scale convolution operator in order to excavate the role of the bone points with different adjacency relations in describing human behavior, and the features obtained by this network are used to action identify. Experiments on MSR-Action3D dataset and HDM05 dataset achieve better recognition rate.

Key words: human action recognition, spatial-temporal feature, deep convolution network, deep learning, action matrix