基于3D卷积神经网络的视频人体行为识别

Video Human Behavior Recognition Based on 3D Convolutional Neural Network

基于视频的行为识别在我们的生活中有着至关重要的作用,比如智能家居、智能安防等。论文提出了一种新的基于3D卷积神经网络的深度学习的视频人体行为识别方法。该神经网络是将VGG-16网络扩展成3D形式,在此基础上加入残差块结构,在增加网络深度的时候能够减少训练误差,经过均值池化层后,输入到LSTM层,识别样本数据集中的各种行为。对比实验结果表明,该方法正确率达到了89.6%。

Video-based behavior recognition plays a vital role in our life,such as intelligent home,intelligent security and other aspects.This paper proposes a new deep learning method of video human behavior recognition based on 3D convolutional neu-ral network.The neural network is an extension of VGG-16 network into 3D form,on this basis,to join the residual block structure,while increasing network depth can reduce the training error,and after a mean pooling layer,input to the LSTM layer,identifica-tion of various behavior of the sample data set.The comparative experimental results show that the accuracy of the method reaches 89.6%.