Research On Activity Recognition Probability Model Of Multi-sensor Time Series

Sensor-based human activity recognition is a kind task of recognition and classification that collects and analyzes the acceleration and angular velocity generated by the subject during the activity,and restores the subject's actual action.It has been widely used in medical care,smart home and other fields.Hidden Markov Model(HMM)and Conditional Random Field(CRF)are probability models that are good at processing sensor time series data,but there are also shortcomings in time series modeling.In view of the following deficiencies,this paper improves the two models and applies them to human activity recognition to verify the performance of the models and improves the accuracy of activity recognition.Firstly,the Baum-Welch algorithm used in HMM training is a hill-climbing algorithm,which is easy to get the local optimal solution.Therefore,the structured Support Vector Machine(SVM)was combined with HMM to construct a structured SVM-HMM framework for human activity recognition,and compared with HMM based on Gaussian Mixed Model(GMM-HMM).Experiments show that structured SVM-HMM can better describe the way of human activities than GMM-HMM.Secondly,the quasi-Newton algorithm BFGS algorithm used in traditional CRF training optimization takes up large memory and slow operation speed.For this problem,The Limited Memory BFGS with Bounds(L-BFGS-B)algorithm was used to optimize the CRF model for the modeling of activity recognition,and compared with traditional CRF,decision tree,logistic regression and other machine learning methods.The results show that the CRF model based on L-BFGS-B algorithm has low computation complexity,low iteration cost and high operation efficiency,and has better recognition effect than other algorithms.Finally,the bidirectional Long Short-Term Memory(LSTM)neural network and CRF were combined to construct the bidirectional LSTM-CRF depth framework for activity classification,and the model was compared with the bidirectional LSTM network and CRF model.Bidirectional LSTM-CRF has both the characteristics of the bidirectional LSTM learning context information and the characteristics of the CRF model consideringthe dependencies between the output label sequences,and the features can be extracted automatically.The results show that the recognition effect of the bidirectional LSTM-CRF model is better than the other two models.