Graphics Reference
In-Depth Information
Table 9.2 Detector performance during testing at Camp Roberts in conjunction with TNT 11-3
Flight
Images
Vehicles
P T
N P
P F
Recall
FPR
2011-05-06
113
461
339
122
291
0.7354
0.2051
2011-05-07
398
1,414
290
1,124
570
0.2051
0.7983
Overall
511
1,875
629
1,246
861
0.4702
0.5017
Table 9.3 Detector speed performance on the ground station (top, GCS), and the embedded PC-104
(bottom)
Flight
) processing
2010-08-08 3,914,757 18,656 504.40 (46.17) 1,494.11 (149.55)
2010-08-11a 4,242,232 4,198 542.70 (122.92) 1,532.57 (156.56)
2010-08-11b 4,165,580 2,394 500.65 (41.36) 1,473.77 (155.13)
2010-08-12 2,532,752 4,589 511.79 (46.04) 1,449.56 (205.83)
Overall GCS 14,855,321 29,839 514.89 (64.12) 1,487.50 (166.77)
2011-05-06 472,171 17,167 548.73 (73.98) 3,020.46 (353.45)
2011-05-07 2,015,450 20,700 667.02 (551.67) 3,093.87 (395.62)
Overall PC-104 2,487,621 37,867 607.87 (312.82) 3,057.16 (374.54)
Image sizes are cumulative for the entire flight and measured in kB. Times are reported in ms
Image size
Cropped size
μ
(
σ
) access
μ
(
σ
from TNT 10-4 (Table 9.1 ) and the first flight from TNT 11-3 were performed at the
same altitude with only small deviations, the second flight from TNT 11-3 exhibited
frequent and pronounced altitude changes. The raw images of the second flight had
more blur, probably partially due to crosswinds and partially due to the rapid position
and attitude changes of the aircraft during image exposure. Additionally, there was no
feedback from the gimbal to the PC-104 so the camera would only take a picture when
the gimbal is stable. Changing altitudes was an attempt to find reduced crosswinds
to stabilize the flight and to reduce the blur. As the UAV flew higher, the vehicle
images were smaller, and smaller vehicles reduce the number of features available
to the detector. This leaves less room for error and causes a reduction in detection
performance.
For the purpose of comparison, the same parameters were used for all test flights.
However, reducing the required number of neighboring detections or scaling the
scanningwindowwith a smaller scale factor (closer to 1) could improve performance.
Naturally, this would increase the total number of windows searched and hence the
processing time. This time might be reduced by accounting for the smaller detection
sizes and reducing the maximum detector scale. The implementation tested here did
not incorporate a feedback for UAV altitude into the detector settings.
Table 9.3 shows the speed performance for embedded vehicle detection. The mean
processing time was 3,057ms, with no large discrepancy between the mean process-
ing time of the two flights. The mean access time was 607
.
87ms.
 
Search WWH ::




Custom Search