Graphics Reference
In-Depth Information
9.5.4.1 Discussion
Embedded processing achieved accuracy performance in the same rough order of
magnitude as GCS processing. The file access time was unchanged, but the image
processing time approximately doubled. The PC-104's reduced on-chip cache size
is a probably culprit for this unsurprising increase. The PC-104 also performed both
flight control and detection.
Reviewing the imagery, there appears to be some halo and blurring around the
edges of the vehicles that was not present to the same extent on the first five flights.
Object boundaries and edges contain a lot of information, and many potential fea-
tures for a haar-like feature-based detector to utilize. This, combined with additional
blurring of the imagery, is most likely the cause of the reduced recall.
9.5.5 Embedded System Operation
The overall system as described in Sect.
9.2.1
worked as desired: the sensors reported
to software processes which were granted sufficient resources by the hardware plat-
form to effectively act as a true real-time system. Combining flight control and
payload processing on the same PC-104 board resulted in space and power sav-
ings without incurring negative effects. The CPU load was measured in two instru-
mented flights. As expected, the image analysis task was demanding on the CPU, but
only intermittently. Presumably, image transfer and load operations alternated with
processing on the CPU, giving other processes sufficient time to run.
9.5.6 Operational Results
The embedded image analysis enabled longer range communication, continued oper-
ation despite severely degraded network performance, more rapid availability of crit-
ical information, and filtering of excessive data. Had the network bandwidth been
sufficient to transmit all images at full resolution, a single human operator likely
would not have been able to thoroughly search a 4,000
3,000 pixel image every
other second for tiny vehicles. A person reviewing these images would have to zoom
in and pan over the 12-megapixel images as each is larger than the screen size. The
repetitious task also bears the danger of operator fatigue and complacency. With-
out embedded analysis, most images would not be available until after landing the
aircraft. With embedded analysis, only the most pertinent information was imme-
diately downloaded, which at the same time reduced the human operator load to a
very manageable level: inspect a few cropped areas every few seconds. Compared to
downloading full-size images, embedded processing yielded a bandwidth reduction
by a factor of 500. The reduced bandwidth and reduced operator load meant that
×
