Locati[o]n-based activity recognition
Automatic recognition of human activities can support many applications, from context aware computing to just-in-time information systems to assistive technology for the disabled. Knowledge of a person's location provides important context information for inferring a person's high-level activities. This dissertation describes the application of machine learning and probabilistic reasoning techniques to recognizing daily activities from location data collected by GPS sensors.In the first part of the dissertation, we present a new framework of activity recognition that builds upon and extends existing research on conditional random fields and relational Markov networks. This framework is able to take into account complex relations between locations, activities, and significant places, as well as high level knowledges such as number of homes and workplaces. By extracting and labeling activities and significant places simultaneously, our approach achieves high accuracy on both extraction and labeling. We present efficient inference algorithms for aggregate features using Fast Fourier Transform or local Markov Chain Monte Carlo within the belief propagation framework, and a novel approach for feature selection and parameter estimation using boosting with virtual evidences.In the second part, we build a hierarchical dynamic Bayesian network model for transportation routines. It can predict a user's destination in real time, infer the user's mode of transportation, and determine when a user has deviated from his ordinary routines. The model encodes general knowledge such as street maps, bus routes, and bus stops, in order to discriminate different transportation modes. Moreover, our system could automatically learn navigation patterns at all levels from raw GPS data, without any manual labeling! We develop an online inference algorithm for the hierarchical transportation model based on the framework of Rao-Blackwellised particle filters, which performs analytical inference both at the low level and at the higher levels of the hierarchy while sampling other variables.