Graphics Reference
In-Depth Information
(a)
(b)
outer box shape
inner box
safe zone
Fig. 4.18 First indoor experiment: a Box as surrogate rooftop; b hand-labeled landing map, outer
box shape is the edge of the box surface, inner box safety zone is the box surface without the 13 cm
border region
conducted with the Asctec Pelican carrying the Asctec Mastermind flight computer.
Even if the whole processing pipeline runs onboard, for the presented experiments,
we recorded the data and processed it offline.
4.4.4.1 Indoor Evaluation
The first indoor experiment was designed to generate quantitative error metrics from
hand-labeled ground truth data. We fly our quadrotor system at three different alti-
tudes over a box (57
27 cm 3 ,Fig. 4.18 ) to simulate a rooftop landing scenario.
In all experiments, the vehicle radius was set to 13 cm to allow for a sufficiently-sized
valid landing area in the middle of the box, and the arbitrary ground-level cut off
threshold was set to 20 cm. For ground truth, the true landing area in the middle of the
box surface was marked by corner marks that were located at 13 cm distance from the
box edges, and these were manually identified in the input images (see Fig. 4.18 b).
The first three rows of Table 4.2 give an overview of the evaluation results. Altitude,
baseline, and feature reprojection error during image alignment correspond to the
average value for each experiment. For this evaluation, we only considered frames
where the box surface was completely visible in the disparity image to avoid bor-
der effects, and were valid stereo results could be calculated (enough baseline and
more than 40 feature matches for image alignment) (“frames visible”). Within these
frames, we defined a successful detection (“frames successful”) as frames where at
least one valid landing location was detected on the box and no landing location was
detected falsely on the ground. For these successful frames, we also calculated false
positives (FP) rates (pixels that were classified as valid landing area that are located
in the border area), false negatives (FN) rates (not as landing area classified pixels
that were located in the correct center area of the box). Note that we only consider
pixels with valid disparity values in this metric.
×
57
×
Search WWH ::




Custom Search