Target Classification using Compressive Sensing Radar without Image Reconstruction
Student researchers: Alex Serrano (EE’16), Eli Friedman (EE’16)
Collaborating Institution: Maxentric Technologies (Hoboken, NJ), co-founded by Kamran Mahbobi (EE’91). Project manager: Brian Choi (EE’10).
Research grant sponsored by U.S. Air Force.
Develop algorithm and processing techniques to permit target detection and classification to be performed directly in the compressed sensing domain, rather than in the reconstructed data domain.
There is a tremendous amount of research being performed in areas related to compressive sensing. This research centers on the development of techniques that allow sparse or compressible signals (not strictly images) to be reconstructed from samples taken at rates significantly below the Nyquist rate. Rather than using a sensor system that requires high-rate sampling followed by extreme data compression, one can develop a sensor system that initially makes low-rate measurements directly in the compressive domain. Under the compressive sensing theory, it is possible to show that under some reasonable assumptions about the sensed signal and the selection of a measurement matrix that satisfies certain reasonable conditions, a signal can be perfectly reconstructed with sparse (well below Nyquist) sampling. A natural progression from the realization that a signal can be reconstructed from these sparse measurements is the realization that it should be possible to perform processing operations on the equivalent representation of the sensor data in the low-dimensional sparse domain directly rather than first having to reconstruct the signal. Obviously if the signal can be reconstructed from the sparse measurements all the information available from a reconstructed signal must be present in the sparse measurement domain. Therefore, it should be possible to perform processing and reasoning techniques directly in the measurement domain without going to the expense of signal reconstruction. There would be at least two advantages to being able to perform signal processing directly in the sparse measurement domain. The most obvious advantage is the elimination of the costs (resources/time) to reconstruct the signal. A second advantage may be a reduction in the raw computational throughput required that results from the significant reduction in the sheer quantity of the raw data (i.e., data quantity commensurate with the sparsity of the measurement space).
Significant advancement in radar based imaging techniques have been made over the last several decades. Once an image is formed, a number of image processing techniques can be applied to classify objects in the field of view. The goal of the joint effort by Maxentric and Cooper Union is to develop a radar based processing system in which compressive sensing techniques are used to acquire a limited set of radar measurements, which would normally be sufficient to reconstruct a high quality image, but to perform target classification directly on the compressed data set, bypassing the need for performing the full image reconstruction.