Home
Speakers
Schedule
Location and Directions
Hotel Information
David Castanon
Boston University
Title: Distributed Coding and Classification using Compressed Sensing
Abstract: Consider the task of hypothesis testing at a fusion center in a sensor network. Source coding before transmission can reduce the communication cost by exploiting redundancy in the data/measurements. Redundancy in the data arises from intra-sensor and inter-sensor correlations. Conventionally, intra-sensor correlations are exploited using transform coding while the inter-sensor correlations can be handled using distributed source coding techniques, such as Slepian-Wolf coding, which require knowledge of the precise correlations during encoding. Compressed sensing is a different paradigm of sensing and compressing a class of signals by taking random projections, which has been successfully applied to exploit the intra-sensor correlation structure. This talk discusses extensions of compressed sensing for networks of distributed sensors that exploit intra and inter-sensor correlations simultaneously. In particular, we discuss different approaches where compression at each sensor is done using only local information, and develop bounds on the relative and communications costs of the different techniques. The results are illustrated in classification of objects with unknown pose using projected image data. Joint work with Ajay Bangla.