摘要： 为将卷积神经网络(CNN)应用到视频理解中,提出一种基于训练图CNN特征的识别算法。利用图像RGB数据识别视频人体动作,使用现有的CNN模型从图像中提取特征,并采用长短记忆单元的递归神经网络进行训练分类,研究CNN模型和隐层的选择、优化、特征矢量化和降维。实验结果表明,与使用图像RGB数据注意力模型的算法和组合长短期记忆模型算法相比,该算法具有更高的准确率。

关键词: 人体动作识别, 深度学习, 卷积神经网络, 递归神经网络, 记忆单元

Abstract: In order to apply Convolutional Neural Network (CNN) to video understanding,a recognition algorithm based on trained image CNN features is proposed.Image RGB data is employed to recognize human action in videos.Off-the-shelf CNN models are used to extract features from images,and classification is made by recurrent neural networks with Long Short-Term Memory (LSTM) unit.The research focuses on the choice of CNN architectures and layers,feature vectorization and dimentionality reduction.Experimental result shows that the algorithm has higher accuracy than attention model algorithm and composite LSTM algorithm using RGB data.

Key words: human action recognition, deep learning, Convolutional Neural Network(CNN), recurrent neural network, memory unit

中图分类号: 

• TP391