Scalable Activity Recognition for Sensor Networks
This project demonstrates that with an appropriate analysis methodology, motion-sensor networks are capable of providing useful contextual information to building services. The methodology we have developed offers scalability and robustness by adopting a probabilistic and hierarchical framework. We call this framework Scalable Activity Recognition for Sensor Networks (SARSEN).
Background & Objective: There is locality in building context. The way hallways and intersections are used is the same in all buildings. The way those larger chunks fit together share commonalities, but also begin to have localized meaning. The overall structure of the building is often unique. If we are going to build robust systems that understand how people use building, then we need to account for these realities. The systems need to understand the building blocks of context. Those building blocks can be built and tested in the lab with some confidence that they will be portable. At the same time, systems need to be easily configured to the unique realities of each building.
Interacting people generate ambiguities. Denser crowds generate deeper ambiguities. Ambiguities can lead to erroneous interpretations if an attempt is made to resolve them too early, with too little information. The SARSEN methodology instead delays resolution as long as possible, making the best local decision possible, but explicitly representing ambiguity that remains, so that it may be resolved at higher levels of context or larger scales of space, where disambiguating context may be available.
Technical Discussion: Empirical results show the methodology to be sound. Results include robust localization of meeting rooms, gathering points, and building resources such as printers. We also show the ability to recover interconnectedness in a building and then track changes in that social network over time using probabilistic techniques to accumulate evidence of behavioral patterns from collections of ambiguous motion traces. These results prove that the SARSEN framework is capable of correctly analyzing the complex behavior in our laboratory despite significant ambiguities. These results are fed by a database of over 30 million motion events collected from hundreds of sensors at MERL over the course of the year. sensors are the product of the Reduced Operating Cost Sensors (ROCkS) projects.
Contact: Christopher R. Wren
Technology Area: Sensor and Data Systems
Modification Date: August 2, 2007
