Research And Application Of Human Activity Recgnition System

Human activity recognition(HAR)systems are widely used in many fields such as personal health management,medical treatment and somatosensory entertainment,and HAR systems can be used as a way to perceive people's current status and thus become a good way of human-computer interaction.Besides the threotical analysis in academic filed,industrial companies has also put huge investment on developing HAR systems.In the current application market,there exist many popular applications such as “gudong”,“Wechat steps” etc.,which also illustrate the great commercial value of HAR.The research on HAR originated from using camera to collect image information and using computer vision technology to conduct behavior recognition.Nowadays we use the wearable device and the built-in sensors in smart phones to collect data for behavioral recognition.In particular,smartphone-baed behavior recognition system is more in line with the user's daily life and no invasion of the user's life.Based on the above research background,this thesis comprehensive studies the problem of HAR,and implements a HAR prototype.The contributions of this thesis mainly include the following three parts:1.In the feature engineering,the three key factors such as the performance of the classification algorithm,the feature reduction effect and the feature cost are comprehensively considered,and the PSO search capability is used to realize feature confidence and feature cost baesd feature selection.In additional,the proposed method has been verified in some UCI public datasets.The experimental result shows the effectiveness of the proposed method.In Lung dataset,compared with the standard BPSO-based feature selection,the proposed method can improve the classification accuracy of KNN by 3.125% and the iteration of particle swarm iteration can be advanced by 4 times.2.Supervisor learning algorithm is the most important part of the HAR system.A good classifier can greatly improve the overall performance of the system.Therefore,the optimization of the algorithm model has also become one of the focuses of our work.The classifier was logically optimized by constructing a hierarchical topology identification scheme using behavioral clustering techniques and verified in the UCI dataset to prove the feasibility of the scheme.In comparison with other popular classifiers such as Random Tree,J48,Bayes Net,KNN and Decision Table,etc.,thorough experiments on the realistic dataset(UCI HAR repository)demonstrate that GCHAR obtains the best classification accuracy,reaching 94.1636%;3.Based on the theoretical results,this thesis designs and implements a complete Android smartphone-based HAR system.The system includes Android APP and Web-side management system.It uses Android built-in sensors to collect human behavior data,and performs energy engineering such as feature engineering and algorithm modeling recognition on the server side.The implementation of this system includes many core parts of the HAR system or basic machine learning project,which has practical engineering significance.