Research On Activity Recognition In Smart Home Based On Conditional Random Fields

Over the last decade there has been a considerable effort from researchers at home and abroad in advancing techniques for activity recognition in ambient intelligence (AMI) environments. However, some existing approaches require specialized (e.g. wearable) sensing devices, which likely cause inconvenience to residents’ lives. Other approaches prefer sensors that neither violate user privacy nor change the habits of residents, but the recognition accuracy of single-resident activity is not high enough, the study on multi-resident activity recognition algorithms is at infant stage, and the study on abnormal activity recognition is still rare. So more and in-depth research into activity recognition in AMI environments is still expected.Conditional Random Field (CRF) is a sequence probabilistic graphical model, first applied to label and segment sequential data. Since CRF-based activity recognition in AMI environments is rare, this thesis proposes to exploit CRF and its extended models to advance activity recognition algorithms in smart homes. And the contributions are as follows:(1) A CRF-based activity recognition framework is presented, and a linear chain Conditional Random Fields (LCRF) based approach is instantiated and further improved through combined features that treat the state changes of several sensors relevant to one activity as one observation feature. Experimental results show that LCRF achieves better recognition accuracy for most activities comparing NB and HMM, and the feature combined method can not only reduce the time for model training and activity recognition, but also reduce redundancy and improve recognition precision.(2) A Latent-Dynamic Conditional Random Field (LDCRF) approach is proposed to model the relationships between activities and their sub-activities. To verify the validity of LDCRF, we use multi-classification metrics to measure results, and compared with the existing classical activity recognition method SVM, HMM, LCRF and related HCRF. The experimental results show that the activity recognition method based on LDCRF is better than other methods.(3) Two-stage Hidden Markov Model (TSM-HMM) and two-stage LCRF (TSM-LCRF) are proposed by defining the combined label and its state set to represent fixed priori knowledge in multi-resident environment. The experiments with multi-resident activity dataset in Washington State University demonstrate that our two-stage method is better than existed multi-resident activity recognition methods.(4) HCRF and LCRF based approaches are proposed to recognize three kinds of abnormal activities which are common in elderly smart home. To verify the effectiveness of the method, we designed three simulation experiments according to "forgotten", "new activity", and "slow activity". The results show that HCRF based method is better than feature vector distance based method in dealing with "forgotten" and "new activity", and LCRF method can also accurately recognize "slow activity".