orb sift surf - copernicus.org · the algorithm is compared to two other feature tracking...
Post on 08-Aug-2018
214 Views
Preview:
TRANSCRIPT
�������������������������������������������
Sea ice drift from SAR using feature trackingStefan Muckenhuber, Anton Korosov and Stein Sandven
contact: stefan.muckenhuber@nersc.no
Nansen Environmental and Remote Sensing Center, Norway
Abstract A computationally efficient, open source feature tracking algorithm, called ORB (Oriented FASTand Rotated BRIEF), is adopted and tuned for sea ice drift retrieval from Sentinel-1 SAR images. The bestsuitable setting and parameter values have been found using four Sentinel-1 image pairs representativeof sea ice conditions between Greenland and Severnaya Zemlya during winter/spring. The performance ofthe algorithm is compared to two other feature tracking algorithms (SIFT and SURF). Applied on a testimage pair acquired over Fram Strait, the tuned ORB algorithm produces the highest number of vectors(6920, SIFT: 1585 and SURF: 518) while being computationally most efficient (66 s, SIFT: 182 s and SURF:99 s using a 2.7 GHz processor with 8 GB memory). For validation purpose, 314 manually drawn vectorshave been compared with the closest calculated vectors, and the resulting root mean square error of icedrift is 563 m. All test image pairs show significantly better performance of the HV channel. On average,around four times more vectors have been found using HV polarisation. All software requirements necessaryfor applying the presented feature tracking algorithm are open source to ensure a free and easy implementation.
Manuscript: http://www.the-cryosphere-discuss.net/tc-2015-215/Sea ice drift algorithm: https://github.com/nansencenter/sea_ice_drift
10 20 30 40 50 60Patch size [pixel]
0
1
2
3
4
5
Matc
hes/
10
km2
Fram StraitSvalbard NorthFranz Josef LandKara SeaHHHVHH+HV
0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20Brightness boundary HH
0
1
2
3
4
5
Matc
hes/
10
km2
Fram StraitSvalbard NorthFranz Josef LandKara SeaHHHV
0.005 0.010 0.015 0.020Brightness boundary HV
2 4 6 8 10 12 14Pyramid levels
0
1
2
3
4
5
Matc
hes/
10
km2
Scale factor = 1.2
Fram StraitSvalbard NorthFranz Josef LandKara SeaHHHVHH+HV
Recommended parameter setMaximum number of keypoints to retain 100 000Resolution reduction during pre-processing 0.5Size of descriptor patch in pixels 34Number of pyramid levels 7Pyramid decimation ratio 1.2Brightness boundaries for HH channel [0,0.08]Brightness boundaries for HV channel [0,0.013]Threshold for Lowe ratio test 0.75
Algorithm tuning
S. Muckenhuber
ORB SIFT SURF
0
5
10
15
20
25
30
35
40
45
> 50
Num
ber
of
vect
ors
0
5
10
15
20
25
30
35
40
45
> 50
Num
ber
of
vect
ors
0
0.3
0.6
0.9
1.2
1.5
1.8
2.1
2.4
2.7
> 3
11
S. Muckenhuber
ORB SIFT SURF
0
5
10
15
20
25
30
35
40
45
> 50
Num
ber
of
vect
ors
0
5
10
15
20
25
30
35
40
45
> 50
Num
ber
of
vect
ors
0
0.3
0.6
0.9
1.2
1.5
1.8
2.1
2.4
2.7
> 3
11
Tuned ORB (first column, 6920 vectors), SIFT (second column, 1585 vectors) and SURF (third column, 518 vectors)Panels: drift vectors (red, first row), number of vectors per grid cell (green, second row) and root mean squaredistance in km (red, third row).
Comparison to SIFT and SURF
Keypoint (red) detection: > 9 contiguous pixels of the surrounding circle(blue) have much lower intensity values than the centreOrientation θ (green): direction to intensity weighted centroidPatch (34x34 pixels): displayed area used for feature descriptionFeature: binary vector from 256 testse.g. (yellow) p(X ) < p(Y ) → τ(p;X, Y ) = 1
1 5 10 15 20 25 30Pixel
1
5
10
15
20
25
30
Pix
el θ
X
Y
Brute Force matching using Hamming distancee.g. ~b1 = 1011101 and ~b2 = 1001001 → d = 2Lowe ratio test: match is accepted, if d1
d2< threshold
ORB algorithm (Rublee et al. 2011)
Manually drawn vectors (white)
30km
CMEMS/DTU pattern matching (blue)
30km
Algorithm Error [m] Compared vectors Average distance [m]ORB vs manual 563 314 1702±1325ORB vs CMEMS 1641 436 2261±1247CMEMS vs manual 1690 201 3440±1105
Validation
top related