SBIR Project Overviews
2004 Phase I SBIR (NAVAIR): Range and Orientation Independent ATR
Automatic target reconition (ATR) is well known to be a very difficut problem. The key result from this Phase I SBIR program was not a new ATR algorithm but a demonstration that a new kind of image (SD image) is much more effective for ATR than the conventional Nyquist image.
Adequately high resolution is essential for recognition by any means. Nyquist images can capture the necessary high-resolution information only by means of a high pixel density, and then are highly pose sensitive because of that high pixel density. SD images, on the other hand, capture high- resolution information no matter how low their pixel density may be.
A recognition experiment was performed in which an "unknown" target image was matched against each of four reference images (one of which was the "true" reference). The ability to recognize the true target was quantified by the difference between the true-match correlation coefficient and the maximum of the three false-match correlation coefficients. In this test the unknown target had a 10- degree pose difference relative to its reference image. It was shown that much better recognition performance was obtained using low pixel density SD images than by convectional (Nyquist) images at any pixel density.
SD images are readily derived from any conventional (Nyquist) image by a simple sub sampling operation. Normalized SD images can also be derived from visible images. They are dimensionless and thus offer the important benefit of being nominally independent of spatial and temporal variations in scene illumination.
This technology is not algorithm specific and is thus potentially applicable over a wide range of ATR systems.
- Much less pose-sensitive
- Many fewer reference images needed (because of their pose insensitivity)
- Much less reference image data (fewer pixels in each reference image)
- Much less processing power required
- High immunity to spatial and temporal variations in scene illumination (with normalized SD images)
It was shown that by the use of SD images, range and orientation independent automatic target recognition (ATR) is feasible with modest processing and memory resources