Algorithm Development - Signature Comparison Analysis


Comparison of Target Signature Synthetic Models to Measured Data

As part of a Phase II SBIR effort performed for the US Army Aviation and Missile Research, Development and Engineering Center (AMRDEC), SDI has developed techniques to quantitatively compare actual target signatures to synthetic models. The purpose of these comparison techniques is to determine if the synthetic models can be used to supplement measured data in weapon system/sensor simulations and evaluations, thereby reducing development costs while maintaining the integrity of the performance evaluation process. The analysis methodology addresses sensors/weapon systems including: Imaging infrared (I2R) sensors/seekers, high range resolution (HRR) real aperture radar (RAR), synthetic aperture radar (SAR), and LADAR. As part of this effort we have developed a Comparison Analysis Tool (CAT) which implements the analysis methodology (see screen shot below).


The comparison approach is based on the fundamental realization that the comparison of synthetic and measured images must be based on how the sensor/weapon system processes those images to extract critical targeting information. Thus, these weapon system processing algorithms or their equivalent must be used as part of the comparison process. The term “images” is used in its most general meaning to include one dimensional, two dimensional and three dimensional spatial data produced by the sensor to represent targets and the background in a scene. Thus, the image could be the one dimensional high range resolution profile from a real aperture radar seeker, the 2D image from an I2R seeker or a 3D LADAR point cloud.

The analysis starts by assembling “balanced” synthetic, {IS}, and measured, {IM}, image sets. (By balanced we mean that the image sets have similar targets at similar poses, ranges, etc.) These image sets are first operated on by algorithms appropriate for the weapon system to extract regions of interest (ROI) that contain both target and background objects. These ROI are then processed by algorithms that represent the additional actual processing performed by the weapon system being evaluated. These processing algorithms can be either tailored, meaning tailored to do just what the weapon system algorithms do, or generic, meaning that they extract the same fundamental information from the images as does the weapon system but are not the same as the weapon system processing algorithms. The output from the processing algorithms is image information, I’, which when formed into a vector may have a very high dimension. For example a correlation filter operating on an image chip (an ROI) will produce a correlation surface with as many values as there are pixels in the image chip. When formed into a vector these correlation surfaces could easily have lengths (dimensions) of a thousand or more. This necessitates the next stage of the analysis, Non-Linear Dimensionality Reduction (NLDR), where the dimension of the information vectors is reduce to a size that can be reasonably handled by a classifier in the next step. The classifier is trained on the reduced dimension information vectors to differentiate between measured and synthetic images, which give the quantitative measure of the similarity that we are looking for.

The classifier confusion matrices are used to identify the percentage of errors that are made when trying to differentiate between synthetic and actual measured image sets. From those confusion matrices we get a single number that ranges from 0 to 100% that captures how well the synthetic and measured image sets match. 100% would be a perfect match and 0% is no match at all. The last step in the analysis is to discover the reasons that the synthetic and measured images don’t match so that the synthetic models may be improved as necessary. For more information contact Steve Smith in SDI’s St. Louis office.

Choose the right for you text message tracker will not be a problem for a long ago time and there were methods to find your particular model is not paying too much or anything track text messages were invented at the same time and everyone liked.