基于循环神经网络的人体行为识别

Research on human behavior recognition based on recurrent neural networks

使用4种类型的循环神经网络模型(RNN、GRU、LSTM、BLSTM)处理手机传感器采集的异构时间序列数据,用于人体行为识别研究.针对4种模型,分别构建自动特征提取方法,并对参数设置进行优化.在公开数据集UCI HAR上进行了行为识别测试实验,实验结果表明,BLSTM模型的识别精度高达95.7%,可以有效地用于行为识别,其识别率和性能优于其他3种循环神经网络,且高于卷积神经网络深度学习方法.

Four recurrent neural network models(RNN,GRU,LSTM and BLSTM)are selected to process heterogeneous time series data collected by mobile phone sensors for human behavior recognition research.For the four models,an automatic feature extraction method is constructed respectively and their parameter settings are optimized.The behavior recognition test experiment is carried out on the public data set of UCI HAR.The experimental results show that the accuracy of BLSTM model is as high as 95.7%,which can be effectively used for behavior recognition.The recognition rate and performance of BLSTM is not only far higher than the other three models,but also higher than the existing deep learning method of the convolutional neural network.