Graphics Reference
In-Depth Information
the expected pixel size of vehicles based on camera parameters and altitude. Still,
as GPS is not without error, and small UAVs are subject to environment and control
system complications, additional experimentation was needed with real-life test data
to select the optimal minimum and maximum scale factors and the step size over
which to move between the factors.)
9.4 Parts-Based Detection
One notices that the different vehicle corners look very similar to each other (see
Fig.
9.5
). To capitalize on this similarity for speed reasons, a separate detector was
built that first detects any of the corners, or “parts,” in any orientation and then looks
for part “constellations” (orientations and locations) that are indicative of an entire
vehicle. (For more details, see Zaborowski [
8
]). Parts-based detection has several
benefits: the parts are faster to detect than the whole object, parts can be detected
even if the object is partially occluded, and multiple parts can be detected in parallel,
possibly even in hardware.
Parts were detectedwith amodifiedViola-Jones method. Instead of binary results,
it scored detections as the number of successfully passed stages divided by the total
number of cascades of the detector.
After nonmaximum suppression, the scored locations of part detections were then
analyzed with a structural model that was trained on pairwise combinations of part
detections. The pair detectors were binary weak classifiers, but with a learned part
score threshold. A second iteration over all training samples determined the predicted
result of every weak classifier for all samples. These predictions were the input to
an AdaBoost algorithm that generated a decision tree.
Figure
9.6
shows detection scores for four parts. The bottompart are the discretized
maps, and highlighted with a pair of circles is one feature of the structural model.
Detection is done in a multiscale scanning window approach.
9.5 Experiments and Results
Several experiments addressed the following questions:
1. Verification: Does the algorithm work as expected, in terms of accuracy and
speed?
2. Which rotation-invariant method works better?
3. What is the usable throughput of the Wave Relay radios?
4. Is the PC-104 board fast enough to keep up with the computer vision tasks along
with the normal aviation workload?
