Recent advances in embedded computing and wireless networking spur the emerging of wireless embedded computing systems that interact with the physical world. This trend has fostered growing effort towards Intelligent Space Systems (iSpaces) that embed intelligence into the physical spaces. One of the most important themes in iSpaces research is to infer high level activity information for the data collected via the wireless embedded computing systems. In this thesis, we will introduce the implement of an indoor iSpaces prototype system, discuss a framework for context-aware computing in iSpaces, and investigate algorithms dedicated to localization, and activity ecognition and analysis in iSpaces. The main work of this dissertation is as follows: (1) Considering the embedded computing and pervasive networking features of iSpaces, we developed wireless networking nodes based on ZigBee protocol; then we configure an wireless network that covers the whole indoor foor of the experimental environment, in which we collected huge number of data about activities about people therein via received signal strength; (2) Mapping the sensor signal into the position of the physical world is one of the fundamental steps towards context-aware computing in iSpaces. Considering the noises and uncertainty of the signal, we apply probabilistic inference methods to formulate the position computing problem into a machine learning problem. Algorithms based on k-Nearest Neighbors method, Multi-class Support Vector Machine method, and Bayesian networks method are applied to validate our methods. Experimental results show high accuracy and precision; (3) Considering the noise of the sensor signals and the uncertainty of user in iSpaces, we apply hierarchical probabilistic graphical method to model the relationship between sensor data, location, action and activity. Than we apply Dynamic Bayesian Networks (DBN) as the inference engine for activity recognition and analysis. Expectation-Maximization (EM) algorithm is adopted for parameter learning in that there are hidden variables in the DBN, and Junction Tree algorithm is applied for DBN inference. To reduce the complexity of parameter learning and inference in DBN, we partition the whole DBN model into two parts. At the low-level a DBN is to inference action sequences from sensor signals. At the high-level two methods are applied for activity recognition from the inferred action sequences. The first method is based on Vector Spa...
修改评论