Graphics Reference
In-Depth Information
Table 9.1
Detector performance on four flights, with true positives
P
T
, false negatives
N
F
,false
positives
P
F
, and the false positive rate (FPR) per image
Flight
Images
Vehicles
P
T
N
P
P
F
Recall
FPR
2010-08-08
714
715
433
282
321
0.6056
0.4496
2010-08-11a
760
95
82
13
69
0.8632
0.0966
2010-08-11b
719
19
16
3
59
0.8421
0.0826
2010-08-12
414
9
7
2
147
0.7778
0.2058
Overall
2,607
838
538
300
596
0.7721
0.2086
with rotated positive samples). Table
9.1
shows the results of processing the images
from four flights on the GCS. All flights are designated by the date they were flown
in year-month-day format. In the case of multiple flights on the same day, the first
flight is denoted as “a” with the second flight on that day denoted as “b”.
Postprocessing combined nearby detections, so any one reported area may be
a collection of multiple detections. Therefore, detection areas containing multiple
vehicles were counted as the number of vehicles they contained. Recall
R
is the ratio
of true positives
P
T
to vehicles present
V
,
R
P
T
V
and was found to be 77.21%.
The average false positive rate (FPR) per image was 0.21. Flights 2010-08-11a and
2010-08-11b had significantly lower false positive rates and higher recall compared
to the other two flights; it is unclear why.
Reading an image fromdisk on the GCS took, on average, 514
=
.
89ms (
±
64
.
12ms),
the vehicle detection algorithm itself about 1,487.50ms (
77ms). Over
14GByte of jpg-compressed images were captured and processed in four flights. The
cropped areas from detections amounted to just under 30 MByte (see also Table
9.3
).
±
166
.
9.5.1.1 Discussion
Although the time of day, above ground level altitude, and detector settings were
constant across all pictures, the UAV itself does not always exhibit consistent flight
and camera platform parameters. For example, the gimble may be in motion while
a picture is being taken, which could result in both blur and a viewing aspect that is
not consistent with the training samples provided to the detector.
The observed recall and FPR are not as good as recent face detectors including
the method this implementation was based on [
6
]. However, recall much better than
chance (realize there are 10,000 s of areas tested with scanning window approach),
and FPR low enough to not swamp network connection. The measured speed showed
a limit of processing an image about every other second.
