Abstract:
Recent research in sensor networks has made it possible to deploy networks of sensors with significant local processing. These sensor networks are revolutionising information collection and processing in many different environments. Often the amount of local data produced by these devices, and their sheer number, makes centralised data processing infeasible. Smart camera networks represent a particular challenge in this regard, partly because of the amount of data produced by each camera, but also because many high level vision algorithms require data from more than one camera. Many distributed algorithms exist that work locally to produce results from a collection of nodes, but as this number grows the algorithm's performance is quickly crippled by the resulting exponential increase in communication overhead. This thesis examines the limits this puts on peer-to-peer cooperation between nodes, and demonstrates how for large networks these can only be circumvented by locally formed organisations of nodes. A local group forming protocol is described that provides a method for nodes to create a bottom-up organisation based purely on local conditions. This allows the formation of a dynamic information network of cooperating nodes, in which a distributed algorithm can organise the communications of its nodes using purely local knowledge to maintain its global network performance.(cont.) Building on recent work using SIFT feature detection, this protocol is demonstrated in a network of smart cameras. Local groups with shared views are established, which allow each camera to locally determine their relative position with others in the network. The result partitions the network into groups of cameras with known visual relationships, which can then be used for further analysis.
Description:
Thesis (Ph. D.)--Massachusetts Institute of Technology, School of Architecture and Planning, Program in Media Arts and Sciences, 2006.Includes bibliographical references (p. 103-111).
