Graphics Reference
In-Depth Information
Fig. 9.5
An example of one positive image rotated in 5
◦
increments
9.3.4 Vision Algorithm Tuning
The hardware specifications including camera optics, gimbal specifications, and
flight altitude were chosen at a sweet spot of resolution on the ground, seamless
area coverage, and motion blur. The core computer vision method placed additional
constraints on the required number of pixels on target, view angle (close to nadir),
and processing time. Sufficiently fast processing speeds and robust accuracy were
achieved through cascade tuning, incremental training with negative images, tuning
of detector parameters on a desktop computer, and scale restrictions at runtime.
Aside fromthe previously discussed orientation experiments, the detectionmethod
was trained with several sets of training data representing different mixes of natural
environments (dirt roads with vegetation around) and city streets (pavement and
buildings nearby). These data were presented to the training algorithm at various
(cascade) stages in an attempt to optimize the speed performance while achieving
the same accuracy. Proper selection of the positive training set improved recall, and
with experimentation it was seen that the negative training set required additional
samples to reduce the false positives. The object size was also varied during training
and subsequent evaluation before settling on 30
30 pixels.
The runtime restrictions amounted to specifying the minimum and maximum
scale values and the size-increment factor between scales for the scanning window
detection approach. This reduced the false positive rate and processing times for each
image. These parameters were set in a configuration file which allowed settings to be
modified on the UAV while it was in flight. (This could be automated by calculating
×
