Research On Human Behavior Recognition Based On RFID And Smart Phone

With the continuous rise of modern science and technology and the changes of the times,radio frequency identification technology(RFID)and smartphones have become an essential part of people’s life.People yearn for a better life more and more strongly,especially in health care,smart home and other aspects.Therefore,human behavior recognition based on RFID and smartphones has become a hot research topic.Simultaneously,the deepening of artificial intelligence theory provides a wealth of knowledge and ideas for human recognition research.Due to the complexity and diversity of human activities in daily life,there is little research on human behavior recognition based on complex actions.The importance of indoor and outdoor human behavior recognition is ignored.This paper mainly studies human behavior recognition based on RFID and smartphone sensor.In order to increase the stability of indoor human behavior recognition,this paper proposes that human behavior recognition based on RFID includes four kinds of daily life behaviors;for outdoor human behavior recognition,this paper proposes human behavior recognition based on smartphone sensor,and proposes a combination of scene and action 19 kinds of daily life actions were identified by machine learning and deep learning.In the design of human behavior recognition based on RFID,considering that single input and single model are not enough to extract features from original data,a manual design of aggregate features combined with automatic feature extraction is proposed.Automatic feature extraction consists of inputting original data into the LSTM model and FCN model,respectively,and manually aggregating features into the DNN model.The three features are fused to construct a multi-input mLSTM-FCN behavior recognition model.In the collection of four kinds of indoor daily life behavior data,through 50%cross validation,the average accuracy rate is 95.56%.In the design of human behavior recognition based on smartphone sensors,multi-scale input is proposed considering that there are various scales of behavior sequence fragments in the original data,too long or too short behavior fragments,and much noise caused by filling and intercepting.Inspired by "video inversion can also represent a picture",reverse sequence input is proposed.Forward sequence and reverse sequence are input into CNN to extract features,and manually aggregate features into DNN model to construct mCNN-DNN model of multi-scale bidirectional sequence.In 19 kinds of data,including scene+action combination,through Five-fold cross-validation,data enhancement and user-defined evaluation index,the score of 0.887 was obtained,and the score of 10 comparative experiments was defeated.This research solves the indoor and outdoor limitations of human behavior recognition,which is necessary to recognize simple and complex actions at the same time and can be widely used in fall monitoring,motion monitoring,smart home and other aspects.