Research On User Daily Behavior Recognition Algorithms For Smart Home

Recently,domestic and foreign researchers have done in-depth research on daily behavior recognition algorithms for smart home,and gained fantastic accomplishments.But the daily behavior recognition algorithms for smart home is less accurate,and the research on abnormal behavior recognition for the elderly living alone is rare.Therefore,it is very meaningful to further study the daily behavior recognition algorithms for smart home.This thesis aims at the current research status mentioned above,it proposes a unified model for user daily behavior recognition and prediction.Then,three types of abnormal behaviors "forgetting","new behavior","slowness of action" of the elderly living alone are studied.Finally,the algorithms are tested and analyzed on experimental platforms.The main work of this thesis is divided into the following sections:(1)For the problem that daily behavior recognition is less accurate,the model for daily behavior recognition based on decision fusion of Cost Sensitive Support Vector Machines(CS-SVM)is established.The simulation results of the improved algorithm are compared with Hidden Markov Model(HMM),Conditional Random Field(CRF)and Support Vector Machines(SVM),which means to verify the former can effectively improve the recognition accuracy of minority classes while ensuring the accuracy of majority classes.Then the Multiple CRF Ensemble Model(MCRF)is selected to predict future behaviors based on typical behavioral sequences,and is compared with HMM,CRF's behavioral prediction effect.(2)Based on the MCRF model,the idea of feature merging is introduced.Firstly,the effectiveness of the feature merging scheme is verified.Based on the WSU Apartment Test Bed,the ADL abnormal database and the database established by Kasteren et al.,three types of typical abnormal behavior recognition for the elderly living alone based on feature merging are performed respectively,and then which are compared with abnormal behavior recognition algorithm based on feature vector distances.(3)The algorithms are tested and analyzed on smart home platforms of Key Laboratory of Industrial Internet of Things & Networked Control and Chongqing Huiju Intelligent Electronics Co.,Ltd.The test results show that the daily behavior recognition scheme of this thesis obtains the highest Recall,Precision,and F-measure,which is 85.7%,86.2%,and 87.4% respectively;the F-measures of the behavioral prediction scheme are increased by 11.78% and 6.53% respectively;the anomaly recognition scheme can reduce the error caused by redundant information,improve the real-time performance of abnormal behavior recognition(total time of measurement is decreased by 18.4s) and accuracy.