TensorFlow平台上基于LSTM神经网络的人体动作分类

Human action classification based on LSTM neural network on TensorFlow

随着人体运动数据采集技术的发展,基于数据的人体运动的研究越来越受到人们的关注。人体运动的研究在医疗康复、运动训练、虚拟现实、以及影视和游戏等领域有着很大的应用空间。人体动作分类就是基于大量已标注动作名称的人体动作,对未标注的人体动作进行分类标注。在本文中,研究提出了一种基于长短时记忆网络(LSTM)的人体动作分类模型。首先,将人体动作表示为时间序列的形式。然后,将人体动作序列逐帧输入到去掉输出层的正向和反向LSTM中,并将隐藏层输出依次送入Mean pooling层和逻辑回归层得到最终的分类结果。最后,研究利用目前流行的深度学习平台Tensor Flow实现本次研发的分类模型并进行训练。基于此,又进一步利用人体动捕数据库HDM05的数据进行实验来验证提出的分类模型,经过训练,该模型在测试集上的分类准确率达到了94.84%。

With the development of human motion data acquisition technology,the research of human motion based on data has attracted more and more attentions. The research of human motion has great application space in medical rehabilitation,sports training,virtual reality,film and television,games and so on. Human action classification aims to classify unlabeled human actions based on a large number of labeled human actions. This paper proposes a human action classification model based on Long Short-Term Memory network（ LSTM）. Firstly,represent human actions as a form of time series; then,input one human action by frame order into two LSTMs without output layer,one is forward LSTM and the other is backward LSTM,and pass the hidden layer outputs of LSTMs into the Mean pooling layer and the logical regression layer to get the final classification results; finally,implement the classification model and train it with the popular deep learning platform of Tensor Flow. The research uses the data of human motion capture database HDM05 to validate the proposed classification model,and the accuracy rate of the classification model reaches 94.84% on test set.