Graphics Reference
In-Depth Information
Fig. 9.6 Graphical illustration of how the four corner detectors are applied to an image
5. Does the algorithm integrate well into the rest of the system? Is the system archi-
tecture (hardware and software) able to produce the expected data, at the expected
quality?
6. Validation: Is the produced information (vehicle detections with the actual, not
ideal, performance) of value to the operators?
After preliminary tests in the lab, all actual flight tests were performed at the Camp
Roberts US Army base in California. Six flights were made in total over five days,
with a gap of several months between the first four and the last two flights. Over 3,100
images were taken that contained over 2,600 depictions of vehicles. Vehicles present
in more than one image were counted each time. Those were manually annotated
with bounding boxes around the vehicles (“ground truth”). A detection was scored
a match only if at least 50% of the respective bounding boxes overlapped.
9.5.1 Vehicle Detection in Aerial Imagery
The goal of this experiment was to demonstrate the feasibility to recognize vehicles
in aerial imagery with a Viola-Jones style detector as described in Sect. 9.3.3 (trained
Search WWH ::




Custom Search