Updates
PublicationsPPAC 2009T. Hartley, A. Fasih, C. Berdanier, F. Ozguner, U. Catalyurek, "Investigating the Use of GPU-Accelerated Nodes for SAR Image Formation", presented to the Workshop on Parallel Programming on Accelerator Clusters (PPAC), held in conjunction with IEEE Cluster'2009, New Orleans, LA, 31 August - 4 September, 2009. Note that Mr Hartley is the author to address correspondences. I would like to thank his advisor, Dr Catalyurek (who also taught my course in emerging computer architectures), for introducing me to such a fine colleague as him. Files: pdf forthcoming. CUDA code for Matlab mex interface and Python (via PyCUDA) forthcoming. Abstract: The computation of an electromagnetic reflectivity image from a set of radar returns is a computationally intensive process. Therefore, the use of high performance computing is required to form images from radar signals in a short time frame. This paper explores the use of distributed memory cluster computers and accelerator technologies such as GPUs for radar signal analysis applications, particularly backprojection image formation. We obtain good results with the use of GPUs and compare their performance in terms of execution time with distributed memory cluster computers. When using a configuration with 4 GPU-accelerated nodes, we achieve speedups up to 3.45x for different input and output data size combinations. Overview: DataCutter Lite is an amazing tool for clusters, heterogeneous or otherwise. PyCUDA is a completely indispensable environment for developing CUDA applications, in Python or otherwise (and it works perfectly in Sage!). SPIE DSS 2009A.R. Fasih, B.D.Rigling, R.L. Moses, "Analysis of target rotation and translation in SAR imagery," Algorithms for Synthetic Aperture Radar Imagery XVI, SPIE Defense and Security Symposium, Orlando, FL, 13 April - 17 April 2009.Note that Dr Rigling is the author to address correspondences. Files: pdf (Copyright 2009 Society of Photo-Optical Instrumentation Engineers. One print or electronic copy may be made for personal use only. Systematic reproduction and distribution, duplication of any material in this paper for a fee or for commercial purposes, or modification of the content of the paper are prohibited. Sigh...) Matlab source code: Point scatterer synthetic aperture radar simulator for the study of randomly vibrating targets, version 1.1. This codebase does not (yet) implement velocity/acceleration/rotation, but can be easily extended to cover this; email me for assistance. Nota bene: the notation and axes orientations in the source code reflect our Asilomar paper (below) and do not match this current paper. Abstract: Synthetic aperture radar (SAR) is a popular tool for long-range imaging of stationary ground objects. Moving targets in the imaged scene will have a mismatch to the matched filter in the image formation process, thus degrading target image quality. In this paper, the impact of uncompensated target motion in SAR imagery is studied in detail. Bounds on allowable target rotation, and random and deterministic translation are derived to maintain image interpretability.  Asilomar 2008A.R. Fasih, E. Ertin, J.N. Ash, R.L. Moses, "SAR focusing performance for moving objects with random motion components," Signals, Systems and Computers, 2008. ACSSC '08. Forty-second Asilomar Conference on, Oct. 2008. Files: pdf.Source code is being prepared for both Matlab and Sage/Python in the interests of reproducible and open research. Please email me if you want me to hurry up and post it. Abstract: Synthetic aperture radar imaging is generally based on the assumption that objects in the scene are stationary. Many approaches have been proposed that compensate for a moving target when one wishes to track or image it. In this paper we consider moving objects that have random fluctuations in their motion, for example caused by engine vibration or road jitter. We study the Cramér-Rao bounds for estimating the location and velocity of a scattering center on such a target, and show that with even small random fluctuations, achievable focusing performance can be several times worse than for a stationary object. Overview: This paper addresses the problem of estimating the location of an ideal point reflector (or "scatterer") in synthetic aperture radar (SAR) data. The accompanying plot is the main (if preliminary) contribution of the paper. It shows two Cramér-Rao lower bounds for estimating the two-dimensional position of a point scatterer vibrating with uncorrelated random motion with varying standard deviation (varying vibration strength), from a SAR at two separate signal-to-noise ratios (SNR---this is the blue vs. green curves). These two lower bounds correspond to the dashed and dashed-dotted curves and give the minimum estimation error achievable by unbiased estimators of two-dimensional location; error in this case is in units of millimeters (lower is better). These two curves diverge because the dash-dotted one uses an approximation that makes it much easier to compute but incorrect for high standard deviations. Note that the horizontal axis is in millimeters of vibration standard deviation, and in the log scale. 1 mm standard deviation you can imagine is a very small vibration, but the plot shows that Cramer-Rao analysis predicts a not inconsequential degradation of between 0.1 mm and 1 mm (hard to see on the log scale but on the order of 2-3 times the error). By the time you get to 10 mm standard deviation, the minimum errors are huge, on the order of 100 times worse than the barely-moving case. This prediction is borne out by the solid curves that represent the error achieved by the maximum likelihood estimator (MLE) of two-dimensional position, which is a brute-force estimator. The paper also briefly considers correlated motion: does estimation accuracy change if the point reflector's motion is low-passed (more smooth than white noise)? The mathematical framework developed allows us to answer this question for low vibration standard deviations (using the linearized approximation) and for a small SAR scenario, we saw that it barely mattered. We did detect something surprising: the more low-pass the motion, the worse the location accuracy. The two ways to understand this is by statistics (N correlated observations of an unknown yields a higher estimate variance than N uncorrelated observations; there's a rho term in there) and by radar imaging (Rigling's work has shown that non-zero coefficients in the Fourier series expansion of the target location engender side-lobes corresponding to those frequencies in the image; assuming fixed vibration power, the lower the motion frequencies, the closer the sidelobes will bunch around the target's mainlobe). There's plenty more work to be done along these lines! SoftwarePGAPhase gradient autofocus (PGA) function for Matlab is here. BackgroundI'm a graduate student at The Ohio State University (official page), working on a PhD in what could be broadly called statistical signal processing (in the electrical engineering department).At this school, a PhD ordinarily consists of three major components, or three different research thrusts that all relate to some central goal. My own research involves synthetic aperture radar (SAR) technology, where basically a radar is mounted on the side of an aircraft which flies around a scene of interest to generate 2d or 3d images of the scene it stares at. This is cool because you can put this aircraft eight miles away from the center of the scene (where maybe some unpleasant things are happening), and it can stare at a circular patch of ground three or four miles in diameter. Manmade objects like buildings and cars show up really well in a SAR image since they're made of metal or concrete, things that reflect electromagnetic energy nicely. (See a large SAR image that OSU folks will recognize.)  My work so far, the first component of my PhD, has been to establish best-case scenarios (Cramer-Rao bounds) for estimating the motion parameters of a moving vehicle from a SAR image. Specifically, my contribution has been to obtain these Cramer-Rao bounds for targets with random motion components (road jitter, engine noise, wind buffeting, etc.) because my advisors and I suspected that it was these tiny random motions that stymied many refocusing algorithms that should have worked really well. We have shown that even a target vibrating with standard deviation of 1 millimeter causes the minimum error in localizing it to double or triple, compared to if it wasn't jittering. I'm also very interested in broadly nonparametric machine learning, meaning mammal-style cognition. I'm deeply moved by Jeff Hawkins and Dileep George's work at Numenta (Hawkins' great talk from a few years ago is online, as is George's dissertation). I hope to contribute to research along those lines and solve problems of interest to the SAR community. The potential for Numenta-style machine learning (completely different than every single algorithm in ordinary "machine learning") is of course vast and is truly the most exciting thing I can imagine working on. Bibliography[1] A. Bovik et al., Handbook of Image and Video Processing. Academic Press, 2005. Available on Google Books. |
