基于循环神经网络的人体动作在线识别方法

Research of Human Action Online Recognition Based on Recurrent Neural Network

由于人体运动的非刚性以及表观特征的多样性,现实场景中对人体动作的准确识别较难。本文提出一种基于循环神经网络的人体动作在线识别方法,把人体关节点作为主要特征,利用深度学习进行动作识别建模。采用循环神经网络学习连续动作的时序关联信息,引入随机权重共享的注意力机制提高训练准确率,避免过拟合现象,实现对人体动作的在线识别。通过UCF数据集进行训练和测试,本文方法达到较高的准确率和稳定性,表明了基于循环神经网络的动作识别模型对现实场景人体动作的在线识别是有效的。

Due to the non-rigidity of human motion and the diversity of apparent features, it is difficult to ac-curately identify human action in real scenes. In this paper, an online recognition method of human action based on recurrent neural network is proposed. The human joint points are used as the main features. The deep learning is used to model the action recognition. The recurrent neural network learns the temporal correlation of continuous actions and realizes online recognition of human ac-tions, in which the attention mechanism is introduced to improve the training accurate rate and decrease over fitting. Through the training and testing of UCF dataset, the proposed method achieves high accuracy and stability, which shows that the action recognition model based on re-current neural network is effective for online recognition of human motion in real scenes.