|Stochastic Systems Group|
Dr. Feng Zhao
Xerox Palo Alto Research Center
Collaborative signal and information processing (CSIP) for distributed sensor networks is an emerging research area, drawing upon traditionally disparate disciplines such as lower-power communication and computation, space-time signal processing, distributed algorithms, adaptive systems, and sensor fusion and decision theory.
Recent advances in wireless networking, microfabrication (e.g. MEMS), and distributed signal processing have enabled a new generation of sensor networks for a range of tracking and identification problems in both civilian and military applications. Examples range from human-aware environments, intelligent transportation grids, factory condition-based monitoring and maintenance, to battlefield situational awareness. However, unlike centralized sensor-poor systems, distributed sensor nets are characterized by limited battery power, frequent node attrition, and variable data and communication quality. To scale up to more realistic tracking and classification applications involving tens of thousands of sensors, heterogeneous sensing modalities, multiple targets, and non-uniform spatio-temporal scales, these systems have to rely primarily on collaboration among distributed sensors to significantly improve tracking accuracy and reduce detection latency.
At Xerox PARC, we have embarked on a set of projects to take a systemic approach to address key CISP issues such as scalable distributed algorithms, progressive accuracy, spatial resolution, and high-level information processing for sensor nets. The key insight is to develop a dynamic feedback mechanism between the high-level structure analysis and node-level signal processing so as to focus the sensing and communication on a when-needed basis. The first problem we are addressing is the combinatorial explosion in data association: assigning signal streams to objects in a distributed setting. I will describe a mechanism we have developed that uses a predictive model to temporally and spatially segment signal streams to drastically reduce the number of possible associations. The second problem we are addressing is the multiple hypothesis management problem in sensor nets. I will describe a method for filtering data and discuss issues concerning data exchange, information utility measure, and data consistency.
Feng Zhao is a Principal Scientist in the Systems and Practices Laboratory at Xerox PARC. Dr. Zhao leads the Collaborative Sensing and Smart Matter Diagnostics Projects that investigate how MEMS sensor and networking technology can change the way we build and interact with physical devices and environments. His research interest includes distributed sensor data processing, diagnostics, qualitative reasoning, and control of dynamical systems.
Dr. Zhao received his PhD in Electrical Engineering and Computer Science from MIT in 1992, where he developed one of the first algorithms for fast N-body computation in three spatial dimensions and phase-space nonlinear control synthesis. From 1992 to 1999, he was Assistant and Associate Professor of Computer and Information Science at Ohio State University. His INSIGHT Group developed the SAL software tool for rapid prototyping of spatio-temporal data analysis applications; the tool is being used by a number of other research groups. Currently, he is also Consulting Associate Professor of Computer Science at Stanford.
Dr. Zhao was National Science Foundation and Office of Naval Research Young Investigators, and an Alfred P. Sloan Research Fellow in Computer Science. He has authored or co-authored over 50 peer-reviewed technical papers in the areas of smart matter, artificial intelligence, nonlinear control, and programming tools.
Problems with this site should be emailed to firstname.lastname@example.org