Information Technology Reference
In-Depth Information
GPU-ESM starts
GPU-SIFT starts
Load one image
Load one image
ESM processing
SIFT processing
No
Tracking fails?
RANSAC solving
Ye s
Store homography
Load homography from SIFT
Fig. 1. Combination of GPU-ESM algorithm and GPU-SIFT
to verify the combination efficiency of both algorithms. 3D region tracking is devel-
oped in the last experiment. Images are captured from a 200 fps camera (Grasshopper
GRAS-03K2M/C). Size is 640
×
480.
4.1
Experiment I: Evaluation with Image Sequence
One image sequence (3000 frames of 640
480 grayscale images ) are loaded into
memory. Then the GPU-ESM and CPU-ESM process the same sequence from the mem-
ory. The number of ESM processing loop is set to 5. The tracking region size is chosen
from 64
×
×
64 to 360
×
360. Their processing speed (fps) and ratio are shown in Fig. 2.
350
35
GPU−ESM(fps)
CPU−ESM(fps)
300
30
250
25
200
20
150
15
100
10
50
5
0
0
0
2
4
6
8
10
12
14
x 10 4
0
2
4
6
8
10
12
14
x 10 4
Pixels
Pixels
Fig. 2. Comparison on processing frame rate of GPU-ESM tracking and CPU-ESM tracking. X
axis is the number of pixels in a square tracking region from 4096 ( 64 × 64 ) to 129600 ( 360 × 360 ) .
Fig. 2 shows that GPU has greatly accelerated the ESM tracking algorithm. Though the
processing speed of both decreases with the increase of tracking region size, GPU-ESM
can still work at a relative high speed. As the 'GPU/CPU Ratio' increases with the pixel
number, it also shows that GPU is more preferable for highly parallel tasks.
4.2
Experiment II: Evaluation with Real Application
Input images are from a 200 fps camera. The captured images are processed simul-
taneously by both GPU-ESM thread and CPU-ESM thread. Images extracted from
Search WWH ::




Custom Search