Research On Human Activity Recognition And Context-awareness In Pervasive Computing Environments

Pervasive Computing is a human-centric computing,and the main purpose of Pervasive Computing is to provide con text-aware services to users in the transparent way.Context awareness is a core feature ofPervasive Computing,and the context-aware system can provide personalized services according to users'activity.Human activity is a main context in Pervasive Computing Environments.and Human Activity Recognition(HAR)is a main research area of Pervasive Computing.Recently,with the development of microelectronics,wearable devices are miniaturized,and have the characteristics of low cost,portable,and high computational power,more and more researchers are focusing on wearable-based HAR.Cross-user activity recognition problem is one of the important issues when research wearable-based HAR,and the performance of cross-user activity recognition is a main aspect of the model's generality.With the development of the Internet of Things,there are a variety ofsensors in Pervasive Computing environments.As context are obtained from different kinds of sensors.the context may be heterogeneous.inaccurate and dynamic.How to effectively model these different types of context is a key issue in Pervasive Computing,a good context-aware model not only express the relationship of contexts and the logic of context-aware service,but also support context reasoning to obtain high-level context.This dissertation was focused on deep learning algorithm for wearable-based HAR,and formal model methodology for context-aware service.The main contents and innovation points are as follows:(1)Based on CAE-ELM.a modified hybrid deep learning model DC-KELM(Deep Convolutional Kernel Extreme Learning Machine)was proposed for wearable-based HAR.DC-KELM used CNN.ML-ELM as feature extractor,and used KELM as classifier.CNN was used to extract local dependency features of raw sensor data.ML-ELM was used to extract more significant features,and KELM was used to get stable classification performance.The comparison experiments were carried out on the public datasets,and the experiment results verified the effectiveness of DC-KELM.(2)In order to use personalized approach to deal with the problem of cross-user activity recognition,a strategy of confidence level was proposed which combined the bayesian posterior probability,uncertainty sampling strategy,and the mean cosine similarity.Based on TransRKELM,a personalized approach US-RKELM(Uncertainty Sampling based Reduced Kernel Extreme Learning Machine)was developed for corss-user activity recognition problem.First,an initial:model based on Reduced Kernel Extreme Learning Machine(RKELM)was trained using the data of the specific known users.Second,the RKELM initial model was tested on new user,and the classification results were transformed into posterior probabilities,through uncertainty sampling strategy and the mean cosine similarity between testing samples,the testing samples which had high confidence level were selected to retrain the initial model using Online Sequential Reduced Kernel Ext re me Learning Machine(OS-RKELM).The comparison experiments were carried out on the public dataset,and the experiment results verified the effectiveness of US-RKELM.(3)Based on UML Activity Diagram and CPN,a hybrid context-aware modeling approach was proposed to model leave home scenario context-aware service.First.UML Activity Diagram was used to describe the leave home scenario context-aware service logic,and then the UML Activity Diagram was mapped into CPN through mapping rules,finally,the correctness of CPN model was verified using CPN Tools.