基于RNN的视频模型构建和动作识别策略研究

Research on Video Model Construction and Action Recognition Strategy based on RNN

递归神经网络(RNN)和长短时记忆(LSTM)在处理顺序多媒体数据方面取得显著成就。因此,提出了一种双向长短时记忆的递归神经网络(DLSTM),该方法结合了卷积神经网络(CNN)和递归神经网络的动作识别新方法。首先,利用CNN提取视频的深度特征,构建视频特征模型,以减少冗余和复杂性。然后,利用递归神经网络学习帧特征之间的序列信息。该方法具有学习长序列的能力,能够在一定的时间间隔内通过分析特征来处理较长的视频。实验结果与现有的方法比较,该方法在动作识别方面有明显完善。

Recursive neural networks (RNN) and LSTM have achieved great success in processing sequential multimedia data. Therefore, this paper proposes a bidirectional long short time memory based DLSTM, combining convolutional neural network (CNN) and recursive neural network. First, the depth features of video were extracted by CNN and the video feature model was constructed to reduce redundancy and complexity. Then, recursive neural network is used to learn the sequence information between frame features. This method has the ability of learning long sequences and can process long video by analyzing features in a certain time interval. Experimental results show that compared with the existing methods, the proposed method has a significant improvement in motion recognition.