existing method #2 - iwahashi lab.tech.nagaokaut.ac.jp/works/iwa/papers/icip07_poster.pdf · takagi...
TRANSCRIPT
WATER LEVEL DETECTIONWATER LEVEL DETECTION FOR FUNCTIONALLYFOR FUNCTIONALLY LAYERED VIDEO CODINGLAYERED VIDEO CODING
ICIP 07 2007.9.17ICIP 07 2007.9.17MP-P3: Scalable Video Coding , September 17, 14:30-17:30
M. IWAHASHI S. UDOMSIRI Y.IMAI S.MURAMATSU*Nagaoka Univ. of Technology, *Niigata University, Japan
http://tech.nagaokaut.ac.jp/index_en.html
Purpose of the ResearchPurpose of the Research
Water Level of Rivers via WEB siteWater Level of ivers via WEB siteMinistry of Land, Infrastructure and Transport Government of Japan
http://www.river.go.jp/
▲ safe▲ caution▲ dangerous× closed
Before…Before… After…After…● by Government● large scale, high cost
● by individual● friendly, ubiquitousg g y
IPネットワークIP network
observation point
requirement to the system1. Transfer Water level regularly2. Video signal when necessaryg y
W t L l D t tiWater Level Detection from videofrom video
“Existing Approaches”Existing Approaches
Existing Method #1
A “board” in the water !
Detect these points by image recognition
W l l i d d
1 Y T k i t l “D l t f t t li id l l i t i i
Water level is detected. Takagi Lab.
1. Y. Takagi, et.al, “Development of a non-contact liquid level measuring system using image processing”, Water science and technology, vol. 37, no.12, pp.381-387, 1998.
2. Y. Takagi, et.al, “Development of a water level measuring system using image processing”, IWA conf. instrumentation, control and automation, pp.309-316, 2001.
Existing Method #2
tical
nt
iatio
n
tect
ntal
line
s
Stain on the wall
Vert
diffe
ren
Det
Hor
izon Water level
Verticaldifferentiation
SubtractionExisting Method #2
* i t i
of framesExisting Method #2
* running water remains* Sensitive to rain drops
N. Tsunashima, M. Shiohara, S. Sasaki, J. Tanahashi, “Water level measurement using image processing”, Information processing society of Japan, Research report, Computer vision and image media, vol.121, no.15, pp.111-117, 2000.
Previously Proposed Methods #1 & #2 by the authors
Verticaldifferentiation
Horizontaldifferentiation
Wavelettransformd e e a o d e e a o a s o
Frame subtraction
ExistingMethod #2
No good
Frame
addition Pretty goodPreviouslyProposed
PreviouslyProposedaddition Method #1 Method #2
1. M. IWAHASHI, “Water Level Detection from Video with FIR filtering”, Sixteenth International Conference on Computer Communications and Networks (ICCCN), Aug. 2007.
2 M IWAHASHI S UDOMSIRI YIMAI S FUKUMA “Water Level Detection for River2. M.IWAHASHI, S.UDOMSIRI, Y.IMAI, S.FUKUMA, Water Level Detection for River Surveillance utilizing JP2K Wavelet Transform”, IEEE Asia Pacific Conference on Circuits and Systems (APCCAS), pp.1766-1769, Oct. 2006.
Proposed Method- abstract -
“Functionally” Layered Coding
CWater levelWater levelComponent 1 2 3
Choos
マイクロビジョン(株)
Component 1 2 3
1st priority ●
se nece
Thumb nailThumb nail
p y
2nd priority ●
essary
non-priority ● ● ●
y compponent
videovideo
t(s)
ウェブドゥ(株)
Previous method #2“recognition”
MPEG“compression”x =recognition
Video Input
“compression”
SynchronousFrame Addition
Motion Compensationd P di ti
Recognition Compression
Frame Addition
WaveletTransform
and Prediction
Discrete CosineTransform
ensor node
Water LevelDetection Encoding Encoding
Se
Priority Layer Non Priority LayerSensor node
Network
DecodingDecoding
Inv.Trans.Inv.Trans. eceivers
Water Level
Inv.Trans.Inv.Trans.
High QualityOutput Video
Low QualityOutput Video
Re
Proposed Methodoposed e odVideo InputS d
TemporalHaar Transform
V deo putSensor node
e
1) frame addition &
2) motion compensation
SpatialHaar Transform
ensor node
) p
temporal Haar trans.
Encoding EncodingEncoding
kSe
3) Wavelet trans. &
4) DCT2nd Priority Layer Non Priority Layer1st Priority Layer
Network
spatial Haar trans.
Water LevelDetection
DecodingDecoding
Inv.Trans.Inv.Trans.
Decoding
Receivers
Detection
Water LevelHigh QualityOutput Video
Low QualityOutput Video
Recognition
Proposed Method- details -
temporal Haar transforminput frames time -->
TemporalHaar Trans.
X(t1) X(t2)
TemporalHaar Trans.
X(t3) X(t4)
TemporalHaar Trans.
X(t5) X(t6)
TemporalHaar Trans.
X(t7) X(t8)
1) priority layer
same effect as the Haar Trans.
L(t1) H(t1)
Haar Trans.
L(t3) H(t3)
Haar Trans.
L(t5) H(t5)
Haar Trans.
L(t7) H(t7)
T l T l
frame addition
TemporalHaar Trans.
LL(t1) LH(t1)
TemporalHaar Trans.
LL(t5) LH(t5)
2) priority layer +
non priority layer
TemporalHaar Trans.
= perfect reconstruction
Haar Trans.
LLL(t1) H(t1) H(t3) LH(t1) LH(t5) H(t5) H(t7)LLH(t1)
non-priority layerpriority layers
“spatial” “spatial” transformWhich band is effective
2HL2LL HL22LL
Which band is effective for water level detection?
Land
Water level
LL HL2LH 2HH LL 1HL2LH 2HH
(a): original video
WaterLH HH
(d): band signals (e): 1st priority layer (g): 2nd priority layer
1LH HH1
temporal transform spatial transform(thumb nail video)
ML estimation
Water level
(b): priority layer(c): non-priority layer (f): discrimination result (h):feature value
level
ML estimation & Water level detectionWater level detection
Spatial ML Temporal
DetectedWaterp
transform LL HL estimationp
transform level
After frame addition LH HH
Feature vector Discrimination Feature value D t ti
Feature vectorland water
result of each line Detection errorDiscrimination
error:land:water
Teacher signals Feature vector
Dimension of the Feature Vector= The number of band signals
b d f ML i i
IdealWater
to be used for ML estimation level
ExperimentalExperimental
resultsresults
Sample #2No.2
(a) original frame (c) after spatial transform(b) after temporal transform
ππ πNo.2 1HH1LH
ω
2ω 2ω
π π00 ω 1ω
2ω
π0
1HL2HH2LH
2HL
1ωπ π00 1ω 1ω π0(a) Land region (b) Water region (c) Difference
Which band should be includedWhich band should be included into the 1st layer?
One dimensional feature vector
30
35
%]
1LL2LL3LL
rror
rror Existing…
discrimination only
20
25
Err
or [%
1LH
3HL
tion e
tion e discrimination only
Proposed
10
15
20
min
atio
n 1HL
1HH2LH2HL3LH
minat
minat Proposed…
discrimination & Data size
5
10
Dis
crim
1 stage2 stage3 stage
1HH
2HH3HH
Discrim
Discrim
00 20 40 60 80 100 120 140
Data Size [KB]
3 stageDD
Data Size [KB]Data sizeData size This criteria is added.
Multi dimensional feature vector
65
70
75
Discrimination Error→ c6 is the minrr
or
rror
55
60
65
Erro
r
→ c6 is the min.(existing)c7
tion e
tion e
45
50
orm
aliz
ed
c2
3
c4 Data Sizec7 is the minm
inat
minat
30
35
40N c3c6
existing
c1proposed
→ c7 is the min.
Discrim
Discrim
25
30
25 30 35 40 45 50 55 60 65 70 75Normalized Data Size
Both of them→ c1 is the min.
DD
Normalized Data Size(proposed)Data sizeData size
Multi dimensional feature vector
sample 2
# b dnormalized normalized normalized
130
%%# bandnormalizeddata size
normalizederror
normalizeddistance
NB NR NDstdev 10.00 10.00 5.79mean 50 00 50 00 50 62
110
120
t3.7
%
32.5
%
mean 50.00 50.00 50.62min 34.54 40.45 40.36max 64.63 70.81 62.10
c1 2HH 38.70 41.95 40.36 ← d
80
90
100
emer
it
Mer
it3
c2 3HH-2HH 39.95 43.71 41.87c3 2LH-2HH 44.70 43.25 43.98c4 2LH-3HH-2HH 45.96 44.85 45.41c5 2LH-3HH 40.54 56.71 49.29
proposed
60
70
80
DeM
c6 2HH-1HH 57.37 40.45 49.64c7 3HH 34.54 63.03 50.82c8 3HH-2HH-1HH 58.62 42.25 51.10c9 3HH-1HH 53 21 50 35 51 80
← existing
50
normalized data size
normalized error
normalized distance
c9 3HH 1HH 53.21 50.35 51.80c10 2LH-2HH-1HH 63.37 41.15 53.43c11 2LH-3HH-1HH 59.21 49.29 54.48c12 2LH-3HH-2HH-1HH 64.63 43.05 54.9113 2LH 39 29 67 13 55 00
proposed/existing
c13 2LH 39.29 67.13 55.00c14 2LH-1HH 57.96 52.01 55.06c15 1HH 51.95 70.81 62.10
Sample #1No.1
75
(a) original frame (b) after temporal transform (c) after spatial transform
130
60
65
70
75
100
110
120
130
No.1
45
50
55
60
rmal
ized
Erro
r
c4 70
80
90
100
30
35
40
No
c2c3c4
c1proposed=existing
50
60
70
normalized normalized normalized
2525 30 35 40 45 50 55 60 65 70 75
Normalized Data Size
p p gdata size error distance
(b) proposed/existing
Sample #3No.3
75 130
N 3
(a) original frame (c) after spatial transform(b) after temporal transform
60
65
70
ror 100
110
120 No.3
45
50
55
orm
aliz
ed E
r
c3c1 70
80
90
30
35
40No
c2
c3
c4proposed
c10existing
50
60
normalized d t i
normalized normalized di t
2525 30 35 40 45 50 55 60 65 70 75
Normalized Data Size(b) proposed/existingdata size error distance
Sample #4No.4
75 130
(a) original frame (c) after spatial transform(b) after temporal transform
60
65
70
or 100
110
120 No.4
45
50
55
orm
aliz
ed E
rr
c2 c3c4
70
80
90
30
35
40No
c1proposed
c5existing 50
60
normalized d
normalized normalized d
2525 30 35 40 45 50 55 60 65 70 75
Normalized Data Size
data size error distance
(b) proposed/existing
ExpectedExpected
resultsresults
What is expected in general ?p g
65
70
75 Proposed method- nearest to the origin.rr
or
rror
55
60
65
d Er
ror Existing method
- lowest vertically
c7
tion e
tion e
1050 −⋅+=
mRNX X45
50
Nor
mal
ized
c2
3
c4Proposed m
erit
- lowest vertically.
minat
minat
}errordetection_data_size,{
,1050
∈
⋅+=
X
NXXσ
30
35
40N c3c6existing
c1proposed Existing
Merit
Dem
Discrim
Discrim
2
22 NBNRND +=25
30
25 30 35 40 45 50 55 60 65 70 75Normalized Data Size
MeritDD
Normalized Data Size
Data sizeData size
Conventional methods ...iti l / i lrecognition only / compression only
75 - recognitionn
50 -
(sacrifice)
%29.2911100 −=⎞⎜⎛ −o
gnitio
proposed●
50 %29.2912
100⎠
⎜⎝
reco
●●
existing25 - compression
(improved)ood←
50 -
25 -
75 - %36.35
221100 +=⎟
⎠⎞
⎜⎝⎛go
good← compression
Proposed method ...U ifi ti f iti & iUnification of recognition & compression
both of75 -n both of
recognition & compression50 -o
gnitio
compression(improved)
122 ⎞⎛proposed
●
50
reco
%23185
1221100 ⎟⎠
⎞⎜⎜⎝
⎛ −−
●●
existing25 -
ood←
%23.18+=
50 -
25 -
75 -
go
good← compression