基于融合注意力机制与CNN-LSTM的人体行为识别算法

Human Activity Recognition Algorithm Based on CNN-LSTM with Attention Mechanism

为解决单一的卷积神经网络(convolutional netral network,CNN)缺乏利用时序信息与单一循环神经网络(recurrent neural network,RNN)对局部信息把握不全问题,提出了融合注意力机制与时空网络的深度学习模型(convolutional neural network-long short term memory network-attention mechanism,CLA-net)的人体行为识别(Human activity recognition,HAR)方法。首先,通过CNN的强学习能力提取局部特征;其次,利用长短时记忆网络(long short term memory,LSTM)提取时序信息;再次,运用注意力机制获取并优化最重要的特征;最后使用softmax分类器对识别结果进行分类。仿真实验结果表明,CLA-net模型在UCI HAR和DaLiAc数据集上的准确率分别达到95.35%、99.43%,F1分别达到95.35%、99.43%,均优于对比实验模型,有效提高了识别精度。

In order to solve the problem that a single convolutional neural network(CNN)lacks the use of sequential information and a single recurrent neural network(RNN)cannot fully grasp the local information,a deep learning model(CLA-net)that integrates attention mechanism and spatiotemporal network was proposed for human activity recognition.First,the local features were extracted by the strong learning ability of CNN.Secondly,the temporal sequence information was extracted by using the long short-term memory network(LSTM).Thirdly,the most important features were obtained and optimized by the attention mechanism,and finally the softmax classifier was used to classify the recognition results.The simulation results show that the accuracy of CLA-NET model on UCI HAR and DaLiAc data sets reaches 95.35%and 99.43%respectively,and the F1value reaches 95.35%and 99.43%respectively.All of them are better than the comparative experimental model proposed,and the recognition accuracy is effectively improved.