Fusion Channels: A Multi-sensor Data Fusion Architecture Bikash Agarwalla, Phillip Hutto, Arnab Paul, Umakishore Ramachandran College of Computing Center for Experimental Research in Computer Systems Georgia Institute of Technology 801 Atlantic Drive NW Atlanta, GA 30332-0280 USA Abstract Due to the falling price and availability of sensors, information capture and processing at a realtime or soft realtime rate is emerging as a dominating application space. This class includes interactive multimedia, robotics, security and surveillance applications and many more. A common denominator of these applications is fusion of data gathered by various sensors and data aggregators. In this paper we propose a Data Fusion architecture, specifically geared toward such multi-sensor data fusion applications and report on the prototype we have built. Our infrastructure provides a programming abstraction that offers programming ease, at the same time provides built-in optimizations that are quite complicated to implement from scratch. We show the ease of programming through two sample applications and also demonstrate through various experiments that our system has low over- head and offers better performance compared to otherwise naively written fusion routines. We also demonstrate improved scalability. 1 Introduction We are interested in an emerging class of applications that involve the capture, interpretation and interactive access to continuously streaming data. We believe that such applications, arising from the confluence of mobility, personal media, interactivity, and the ubiquity of small, high-powered computational devices (e.g. sensors), place