Research On Human Activity Anomaly Detection Based On Multi-Sensor

Human activity recognition(HAR)is a popular research field in artificial intelligence,and smartphone-based human activity recognition technologies gain increasing attention by researchers with the rapid development of sensor technologies and intelligent devices.Based on the state-of-the-art,this thesis proposes using a variety of sensor data in smartphone to conduct human activity recognition research,and uses collected data to verify the performance of the proposed model.Firstly,this thesis introduces the research background,current situation,existing problems and key technologies of human activity recognition which based on sensors.The preliminary work are explained one by one to pave way for the research content of this thesis.Secondly,this thesis studies the identification of human daily activities.The sensor data used in this thesis includes the acceleration,angular velocity and linear acceleration collected by smartphone.To minimize the impact of azimuth change on the algorithm,this thesis introduces the amplitude of the sensor signal.Based on the problem of human activity feature representation in current researches and advantages of Convolutional Neural Network(CNN),the thesis utilizes CNN to extract sensor data features.Taking the temporal characteristics and dimensional correlation of sensor signals into consideration,the convolution and pooling adopt one-dimensional convolution and pooling.After the first convolution layer and pooling layer,this thesis combines the data features of the same dimension among different sensor data,and then extracts the deep features through second convolution layer and pooling layer.On the choice of classification algorithm,this thesis selects the random forest which has superior performance.Using the sensor features extracted from previous CNN model,the random forest algorithm can identify most of the human daily activities in this thesis.Thirdly,this thesis studies the detection of human abnormal activity.For fall activity with low frequency of occurrence in daily life,this thesis considers the unbalanced characteristics of human activity samples and studies the characteristics of sensor signal for fall activity in detail,and proposes a dual detection method based on One Class Support Vector Machine(OCSVM)and threshold.In feature extraction stage,the anomaly detection model still uses the constructed CNN model in Chapter Three,and uses the OCSVM algorithm to initially detect the suspected falls,and then uses the threshold method to perform secondary detection considering the change in the human posture of the fall activity.The experimental results illustrate the anomaly detection model presented in this thesis has a very high detection rate for the fall activity.Finally,this thesis summarizes the current work and drafts plans for the future works.