Information Technology Reference
In-Depth Information
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
2
4
6
8
10
12
time(s)
Fig. 5. ZNCC values respecting to time
t=0.00s
t=4.82s
t=4.90s
t=6.62s
t=6.65s
Fig. 6. Combination evaluation. Occlusion happens at t = 4 . 82 s and t = 6 . 62 s . After occlusion is
moved, it can continue tracking (see t = 4 . 90 s , t = 6 . 65 s ). The red boxes shown in the first row
show that by loading from SIFT, the ESM tracking can continue working even with occlusion.
happens at t
6 . 62 s , the ZNCC value fell down (in Fig. 5). The GPU-
ESM detected the tracking error and loaded the homography from GPU-SIFT. There-
fore it can continue tracking at an acceptable accuracy. After occlusion is moved, the
GPU-ESM can continue tracking (see the image sequences in Fig. 6). This has verified
the effectiveness of our combination model.
=
4 . 82 s and t
=
4.4
Experiment IV: 3D Object Tracking
This experiment is to evaluate the 3D object tracking. Thanks to the GPU speedup, we
extend the 2D planar tracking to 3D region tracking based on multiple planes tracking.
In many applications a 3D tracking region can be separated into multiple adjacent planar
regions. We can carry out ESM tracking on each planar region and merge the warped
regions again to realize the tracking task. As shown in Fig. 7, for tracking area of 240
×
416 with two planar regions, the processing speed is 130 fps.
In our previous work of 3D region tracking, the boundaries of template region are
manually chosen. Now the GPU-SIFT is also extend to 3D tracking. A 3D template
region is chosen from current image (two adjacent regions in Fig. 7). Then the object
is moved to a random initial pose(see the warped image of t
0 . 00 s ).GPU-ESM will
continue loading the homographies for two regions from SIFT and tracking the 3D
region on the moving object until t
=
=
1 . 12 s , when the GPU-ESM has found an accurate
Search WWH ::




Custom Search