基于改进长效递归卷积网络的行为识别算法

Action recognition based on improved long-term recurrent convolution network

为充分提取视频序列中人体行为的静态特征与时域特征,提高人体行为识别算法的准确率,结合深度卷积神经网络与递归神经网络,提出一种端到端的网络模型,分别使用多帧叠加的RGB图像与光流图像作为网络输入,将基于RGB图像的人体行为特征与基于光流图像的人体行为特征进行加权融合,作为最终的人体行为特征。实验结果表明,该算法可以有效提高行为识别准确率,在公开数据集UCF101上取得了84.68%的平均准确率,高于改进前长效递归卷积神经网络(82.34%)。

To fully extract the spatial feature and time domain feature of human activity in video sequences and improve the accuracy of human action recognition algorithm,an end-to-end network combining with deep convolution neural network and recurrent neural network was presented.The stacked RGB images and the stacked optical flow images were respectively used as the network input,and the features based on the RGB images and the features based on the optical flow images were weightedly integrated as the ultimate human activity features.Experimental results show that the proposed algorithm can effectively improve the accuracy of action recognition,and obtain the average accuracy rate of 84.68%in the open dataset UCF101,which is higher than that of the long recurrent convolution network（82.34%）.