Graphics Reference
In-Depth Information
Training a detector with the additional samples (14,400 positive samples in 72
orientations versus the 200 positive samples used for the aligned detector) results in a
modest increase in off-line training time. This training technique produces classifiers
with more features, and therefore requires more CPU cycles to process an image.
However, the dashed ROC curves in Fig. 9.7 indicate that the recall across the range
of in-plane rotations was consistently higher compared to that of the aligned training
set. The rotated training set enabled building a rotation-invariant detector, meaning
it can find vehicles in all orientations in only one pass, thereby eliminating the need
to perform multiple passes over incrementally rotated test images. Given the com-
putational runtime expense of image rotation, the possible introduction of artifacts
caused by the interpolation method, and the increased pixel count from the additional
black areas (see Fig. 9.4 ), the rotated training set yielded the preferred method.
9.5.3 Bandwidth Performance of Wave Relay
The available bandwidth between the Rascal UAV and the GCS over Wave Relay
varies with the number of network users, with vehicle distance, altitude, and the
exact relative orientation of the (omnidirectional) antenna. In tests with few users
the throughput for a realistic data set (sending a large, JPG-compressed images to
the GCS) was determined to be around 200kB/s typically. Throughput decreased
roughly linearly with altitude. At that rate, a 12 MPixel image, compressed to 4MB,
took about 20 s to download.
9.5.3.1 Discussion
For the desired coverage and ground resolution, given jitter and vehicle speed, the
sweet spot for vehicle altitude and focal length identified 0.5Hz as the necessary
shooting speed. The download speed therefore is one order of magnitude too slow
to keep up with a high-resolution image every other second.
9.5.4 From GCS to Embedded Processing
After profiling and parameter evaluationwith existing systems, and after procurement
of appropriate embedded hardware, the performance of the algorithms was evaluated
on the embedded hardware during a live flight. Between experiments, a new detector
had been trained with an increased negative training set to reduce false detections
that frequently occurred on the runway and other rectangular objects. Both detectors
had been trained with rotated imagery.
Table 9.2 gives the detection results. The average recall was 47.02% at a FPR of
50.17%. There were large differences between the two flights. While the four flights
Search WWH ::




Custom Search