Abstract
As sensors that measure daily human activity become increasingly affordable and ubiquitous, there is a corresponding need for algorithms that unearth useful information from the resulting sensor observations. Many of these sensors record a time series of counts reflecting two behaviors: 1) the underlying hourly, daily, and weekly rhythms of natural human activity, and 2) bursty periods of unusual behavior. This dissertation explores a probabilistic framework for human-generated count data that (a) models the underlying recurrent patterns and (b) simultaneously separates and characterizes unusual activity via a Poisson-Markov model. The problems of event detection and characterization using real world, noisy sensor data with significant portions of data missing and corrupted measurements due to sensor failure are investigated. The framework is extended in order to perform higher level inferences, such as linking event models in a multi-sensor building occupancy model, and incorporating the occupancy measurement from loop detectors (in addition to the count measurement) to apply the model to problems in transportation research.