Images courtesy SRC Computers
Figure 1: In the Implicit+Explicit Architecture, Dense Logic Devices (DLDs) encompass a family of components that includes microprocessors,
digital signal processors, as well as some ASICs. These processing elements are all implicitly controlled and typically are made up of fixed logic that is not altered by the user.
Figure 2: Systems can be built with a single MAP processor
and microprocessor combination, or when more flexibility is
desired, Multi-Ported Common Memory accommodating up to
three MAP processors and Hi-Bar switches accommodating
thousands of MAP processors can be employed.
Figure 3: SRC servers that use the Hi-Bar crossbar switch interconnect can incorporate common memory nodes in addition to microprocessor and MAP nodes. Each of these common memory nodes contains an intelligent DMA controller and up to 16 GBs of DDR-2 SDRAM.
Figure 4: The MAP processor used in this system was the most powerful SRC-6 MAP processor ever produced. It was coupled to an Intel Pentium microprocessor and used a Fedora Linux operating system.
Figure 5: The second airborne system in production is a 10-module system designed for payload bay 3 of the General Atomics Sky Warrior, but is also usable in other larger manned and unmanned platforms. It contains a dual Xeon motherboard, a Hi-Bar switch, 750 Gbytes of removable encrypted storage, 28 VDC power system, thermal solution and a mixture of up to 10 MAP processors or common memory
Figure 6: This system is being designed to withstand an operating range from –50C to +50C, an altitude limit in excess of 25,000 feet. and will meet shock and
vibration requirements for single engine aircraft weighing less than 12,500 pounds.
Figure 7: A grayscale pixel’s intensity is simply the pixel’s eight-bit numeric value, but the intensity information is distributed among the individual RGB values for a color pixel. To obtain the intensity value
for an RGB pixel, each 24 bit RGB value is transformed from the RGB color space to the Hue-Saturation-Intensity (HSI) color space. The intensity values for all pixels in both frames are then histogrammed.
From these two intensity histograms, a statistical Cumulative Distribution Function (CDF) is created and then normalized for each frame. A mapping function is created from these two normalized CDF
arrays to map the original color pixel intensity values to a new intensity value such that the new intensity value distribution matches the GS pixel intensity value distribution. The original intensity values are
re-mapped and the new HSI image is transformed back into the RGB color space.
Figure 8: The MAP processor’s GCM Bank 0 acts as a frame buffer for the RGB image and GCM Bank 1 acts as a frame buffer for the GS image. In stage
0, two RGB and six GS pixel intensities are histogrammed in parallel every clock. The integer RGB intensity calculation is part of the RGB histogramming
pipeline. After all pixel intensities for both frames are histogrammed, stage 1 calculates the CDF arrays for both histograms for all histogram bins in parallel.
Stage 2 normalizes both CDF arrays in parallel, a single precision floating point (SPFP) calculation. Stage 3 uses both normalized CDF arrays to generate
the histogram matching MAP array. Finally, stage 4 re-reads the RGB image data two RGB pixels per clock from GCM Bank 0 and calculates the HSI pixel
values. The two integer intensity values select two new intensity values from the Map array (generated in stage 3). The two new intensity values are cast to
SPFP, and together with the two SPFP pixel hue and saturation values, are converted back to the 24 bpp RGB color space and stored in GCM Bank 1.
Figure 9: The CPU normalized cross-correlation application is a single threaded, serial implementation of the algorithm shown in Figure 8.
One popular algorithm is the pyramidal version of the Kanade-Lucas-Tomasi (KLT) feature tracker. This tracker works by minimizing the Euclidian distance between two frame’s subimages by an iterative gradient search. The algorithm is computationally efficient and well suited to real-time applications.
The computational kernel of the Kanade-Lucas-Tomasi tracker is an iterative horizontal and vertical convolution operation performed on two successive frames. On the MAP processor, both convolution operations are fully pipelined and executed in parallel at eight pixels per clock.
Image cross-correlation is used to locate and track a template image within a larger image.
An image template is placed on a pixel in the target image and the subimage under the template is compared to the template image for all pixels in the subimage. Normalized cross-correlation is used to reduce brightness variance of the target image and template caused by lightning and exposure conditions. Normalization is performed every comparison step by subtracting the template and subimage mean and dividing by the product of the standard deviations.
What follows is a description of a spatial domain normalized cross-correlation implementation on the SRC-7 Series H MAP processor and compares its performance to a CPU implementation. Three template images (8x3, 8x6, and 8x9) are correlated to image sizes corresponding to each of the six sensor geometries. Since the template images are constant, their calculations are performed once at initialization time and stored on the CPU and MAP implementations.