Information Technology Reference
In-Depth Information
Due to the small eigenvalues of
A
(
x
0
)
(near 0), above approximation can work well
without loosing accuracy.
6) Correlation. This kernel calculates the correlation of warped region
I
and template
region
I
∗
(
m
pixels) with the Zero mean Normalized Cross Correlation (ZNCC):
∑
m
I
∗
(
k
=
1
(
I
(
k
)
−
I
)(
k
)
−
I
∗
)
(19)
2
2
m
k
m
k
I
∗
(
I
∗
)
∑
(
I
(
k
)
−
I
)
∑
(
k
)
−
=
1
=
1
where
I
and
I
∗
are the mean intensity values of warped region
I
and template region
I
∗
,
respectively. As a quality evaluation criterion, the correlation has played two important
roles in our application. On one hand, if the correlation is smaller than a preset lower
threshold, it will be treated as ESM tracking failure has happened. On the other hand,
if the correlation is larger than a preset upper threshold, the iterative ESM processing
loop will stop and continue to process the next input image. As an iterative minimization
method, such threshold to stop the ESM loop is necessary.
3.3 Implementation of GPU-SIFT
In our GPU-SIFT, we transfer Changchang Wu's GPU-SFIT matching[19] to match the
features. Then we adopt the RANSAC method to improve the homography accuracy.
RANSAC method has shown a better performance than least squares methods as it can
effectively remove some of the mismatched pairs of points in GPU-SIFT.
3.4 Combination Strategy
As mentioned above, ESM tracking algorithm can provide a fast and accurate homog-
raphy solution when the solution is near the global minimum point, but its convergence
region is small. For large image difference it will loose tracking. Meanwhile, SIFT al-
gorithm can offer a robust solution in a large region. But it is not fast enough for a real
time visual servo system. Limited by the mismatched outliers, the homography solution
is not so accurate as that from ESM tracking algorithm.
Therefore we combine the GPU-ESM tracking and GPU-SIFT methods to enhance
the system performance. The combination model is shown in Fig.1. Both GPU-ESM
tracking and GPU-SIFT run on GPUs simultaneously to process the input images.
Though the two threads might process two different frames because of their differ-
ent processing speed, the system can still work well because there is no large image
difference between the two images in such a small delay time.
In GPU-ESM tracking algorithm, the ZNCC correlation value will be checked af-
ter processing each image to determine whether ESM tracking failure has happened or
not. If tracking failure happens, GPU-ESM will automatically load current homography
from GPU-SIFT and set it as the new initial value. By this means, the GPU-ESM track-
ing algorithm can continue working. Therefore, the whole homography-based visual
servo system can work smoothly with high reliability at a high processing speed.
4
Experiments
Four experiments have been carried out to evaluate the system. The first two experi-
ments are to evaluate the efficiency of our GPU-ESM tracking algorithm. The third is
Search WWH ::
Custom Search