Research On Activity Recognition Based On Convolutional Neural Networks And Recurrent Neural Networks

The activity data of the human body can be collected autonomously by sensors in some collection devices such as smart phones,wristbands and the like.Analysis of the types of activities contained in these data has important practical implications.The traditional manual extraction feature activity recognition method does not have good generalization ability,and the extraction feature is highly correlated with domain knowledge,which has considerable limitations.At present,there are more and more researches on convolutional networks and recurrent networks in deep learning in the fields of image,voice,translation,etc.However,there are not many related researches in the field of activity recognition,and there are few studies on the application of these network architectures purely for raw data.In view of this,this paper carries out related work on human activity recognition research from the following three aspects.Firstly,the convolutional neural network and the circulatory neural network are explored on the issue of human activity recognition.Specifically,three typical convolutional neural networks and three typical recurrent neural networks were selected for experiments.The six network models were placed on three public data sets for experiments.The data set also experimented with four traditional machine learning methods.Here,a more detailed comparison can be made.Secondly,for the convolution model and the recurrent model,a CRNN model is proposed,which combines the characteristics of the convolutional network and the recurrent network into two forms:joint and discrete.Experiments were performed on the same three data sets for these two different forms of models.Finally,for the feature extraction process of sensor data,a variable-dimensional convolution model different from the fixed convolution dimension in general convolutional network model is proposed.The model combines one-dimensional convolution feature extraction process and two-dimensional convolution feature extraction process.The advantage is to form a variable-dimensional convolution model,which achieves 95.62%and 95.16%classification accuracy on two public data sets,respectively,which is better than the data in published results in recent years.