Dissertation Defense WHAT: Jacky Mallett: "Eye Society: Building Intelligent Co-operative Networks of Smart Cameras" WHEN: Wednesday August 31, 2005, 1:00 PM EST WHERE: Bartos Theatre, MIT Media Lab (E15) DISSERTATION COMMITTEE: V. Michael Bove, Jr. Principal Research Scientist MIT Media Laboratory Ted Selker Benesse Career Development Professor of Media Arts and Sciences MIT Media Laboratory Wayne Wolf Professor of Electrical Engineering Princeton University ABSTRACT: It has recently become possible to build and deploy large quantities of cheap networked sensors with significant local processing capability. These autonomous sensor networks are expected to revolutionize information collection and processing in the environments in which they will be deployed. However, in many cases, both the amount of local data produced by these devices and their sheer number makes centralized data collection infeasible. Camera nodes represent a particular challenge in this context, both because of the large amount of data they produce and also the utility of information from multiple viewpoints in a number of image analysis problems. In order to extract information effectively from these networks, local algorithms are needed which allow nodes to identify those of their peers with useful information and efficiently exchange data with them. In this thesis, we construct a purely local group forming protocol, where each node independently decides whether or not to join a task group based on broadcasts of group membership requirements. We show that this protocol can be used to allow nodes to form a dynamic information network, allowing nodes to cooperate with each other on tasks without identical views of group membership, thus avoiding the problem of maintaining consistency in group membership at each node, since enforcing this across a network is probably impossible. This protocol is used in a network of smart cameras to set up task groups with shared views, allowing each camera to locally determine their relative position with others in the network, and which also thereby partitions the network into groups of cameras with known visual relationships, which can then be used for further analysis. MIT Media Laboratory Home Page | Events Main Index