Research On Human Behavior Recognition Algorithm Based On Smart Phone Sensors

Human behavior recognition refers to the process of recognizing different action categories and behavior patterns of human beings.This process uses computer technology to analyze human behavior data,and finally uses natural language to express.With the rapid development of The Times,smart phones are ubiquitous and indispensable in people’s life.At the same time,sensors embedded in smartphones provide a convenient way to collect data on human behavior.Therefore,the construction of human behavior recognition framework based on smart phone sensors has become a new research direction in recent years,and has been widely used in health monitoring,human-computer interaction,information security and other fields.The establishment of human behavior recognition framework based on traditional machine learning model and neural network model has been a hot research topic in recent years,but there are still many shortcomings in the recognition and classification of human behavior.For example,the current research only focuses on the recognition of specific behavior categories,but lacks the recognition of behavior transformation.In the aspect of feature extraction,there is a lack of attention to the Angle feature in behavioral data.The accuracy of the model used to identify human behavior needs to be improved.In view of the above shortcomings in the current research,this paper proposes three improvements:(1)Increased the recognition and classification of behavior switching actions.The human behavior data set of the latest version of UCI database in 2015 was used to train the classification model.This data set not only contains six specific categories of human behavior,but also provides sensor data of six kinds of behavior transformation from standing to lying down,from lying to standing,from sitting to lying down,from lying to sitting,from sitting to standing,and from standing to sitting.(2)Features describing the rotation amplitude of human body in data samples are added during feature extraction.Are the body acceleration and angular velocity and the Angle between their first difference and the gravitational acceleration,which is the Angle between the x axis,y axis and z axis respectively.Focusing on Angle characteristics can better describe the rotation amplitude of human body in different activities,so as to distinguish different types of human behavior more accurately.(3)Traditional machine learning model and neural network model are trained on WISDM dataset and UCI dataset respectively.Considering the classification accuracy and training time of the model,the support vector machine model of the traditional machine learning model has good classification performance,and the F1 scores of the two data sets are as high as 94.41% and93.16%.For the neural network model,the CNN-Bi LSTM model has better performance,and has achieved good classification effect on both WISDM dataset and UCI dataset,with F1 scores of 95.83% and 93.63%,respectively.