Research Of Human Behavior Recognition Based On Smart Watch

With the rapid development of economy, and the advance of human living condition, more and more chronic diseases appear. Some early symptoms of chronic diseases may appear, which will cause some abnormal behavior.Discovering the abnormal behavior early and give some intervention timely is very important to the patients. The cause of chronic disease is very complex, obesity has been recognized as a main factor of many chronic diseases. In addition to unhealthy eating habits, many obesity is caused by sedentary. So, do some exercise properly is necessary, which can help keep sober mind and prevent becoming fat, and then prevent the happening of chronic diseases. Human behavior recognition technology is the foundation of the human-computer interaction, with the help of computer technology to analyze human behavior. Recognition of human behavior and giving timely reminder can help people realize their own motion and do proper exercise to keep fit. Recognizing human behaviors and keeping track of motion trajectory can effectively know about people's behavior, which is of great significance to predict the occurrence of some chronic diseases.Most of the existing human behavior recognition system is unfit for human daily use, using smart watches for human behavior recognition has better applicability. This thesis defined five kinds of human movement behavior: stay still,walking, jogging, upstairs and downstairs, and used Apple Watch as the platform to study behavior recognition algorithm. Firstly, here designed and implemented a data acquisition system on Apple Watch, and collected the data of human movement from the three-axis acceleration sensor of five behavior. In order to reduce the influence of noise data which caused by unrelated motion in the process of collecting data, this thesis used the median filter algorithm to deal with the original acceleration signal. Considering the collected triaxial acceleration data of five kinds of behavior changing differently and people wearing watch with different habits, this thesis used a feature extraction method based on principal component analysis(PCA)to extract features from triaxial acceleration data from both hands, and finally got ahigher recognizable feature set. Classification algorithm is the essence of behavior recognition algorithm, here we used three classic classification algorithm: Decision tree algorithm, Naive Bayes algorithm, and Neural network classification algorithm to test the feature set above.Experiments proved that the Neural network classification algorithm has a better results. The experimental results show that, recognizing human behavior based on the smart watch has good availability.