Human activity monitoring and modeling at different spatiotemporal resolutions using wireless sensor networks.

摘要

This thesis addresses the fundamental problems and challenges of human activity monitoring and modeling in the context of sensor networks and examines their architectural implications through the design and implementation of BScope, a run-time framework for studying and interpreting human behaviors and activities using distributed wireless sensor networks. Our approach is based on the fundamental observation that human activities are sequences of very primitive actions that take place over space and time. Multiple activities can be described by simply combining these primitive actions in different ways. The role of the sensor network is to continuously monitor a person's location and interaction with different objects over space and time to provide a stream of basic sensing features. When there is information in advance about the type of activities taking place, then recognizing these activities can be seen as a sequential pattern recognition problem. The proposed method suggests to parse the sequence of detected sensing features into higher level human activities in a hierarchical bottom-up processing model that is similar to natural language processing. In essence, we combine the low-level sensors of the network to develop human activity languages. The set of sensing features becomes the human activity alphabet. In the same sense we combine letters to form words, words to form sentences and sentences to form paragraphs, we combine recorded sensing features to describe primitive actions, primitive actions to describe activities and activities to describe macroscale behaviors. When there is no information in advance about the activities taking place, we have devised a methodology for automatically extracting activity information from sensor data streams in a data-driven way. We do so by properly mining spatial, temporal and frequency information from the sensor data stream. In both cases, our system architecture has been designed so that it can concurrently support multiple time scales enabling human activity recognition and modeling at different spatiotemporal resolutions. BScope's ability to efficiently perform human activity recognition and modeling is demonstrated using a two-month dataset recorded by two multimodal home sensor network deployments where two elder persons living alone were monitored.