Upload
jonah-houston
View
215
Download
0
Embed Size (px)
Citation preview
1
Detection, tracking and sizing of Detection, tracking and sizing of fish of in data from DIDSON fish of in data from DIDSON multibeam sonars multibeam sonars
Helge Balk1, Torfinn Lindem1, Jan Kubečka2
1 Department of Physics, University of Oslo, PO.Box.1048. Blindern, NO-0317 Oslo, Norway email: [email protected], [email protected]
2 Biology Centre of Czech Academy of Sciences, Institute of Hydrobiology, Na sadkach 7, CZ 37005 Ceske Budejovice, Czech Republic. e-mail: [email protected],
2
CFD AND CFD AND DIDSONDIDSON
tracking
3D approach
Detection methods
Echogram approach
Introduction
Conclusion
Inc.Video methods
Our main interest Our main interest
As usual to find out abot the fish
How many How big What are they doing
4
Equipment that may be usedEquipment that may be used
Resons-SeabatCoda Octopus EchoscopeDIDSON
Simrad MS70Simrad SM2000 Split beam
DIDSONDIDSON
Dual frequency Identification SONnar Developed for military underwater tasks like
diver night vision and mine searching
Become popular for fish studies Identification ability Can see pictures of the fish. Fish size from geometry, not from TS
6
Our aimOur aim
Develoop a target detector for DIDSON data
Can vi use the Cross Filter Detector CFD develooped for ordinary echogram
If not, can we optimise it to fit the DIDSON data
Or is there something to learn from the video world
7
DIDSON problemsDIDSON problems
Low snr,
Low dynamic span,
Not calibrated,
Not veldefined sample volume
Only x,z, but no y position information
9
12
CFD AND CFD AND DIDSONDIDSON
Tracking
Echogram approach3D approach
Detection methods
Aim, material and methods
Introduction
Conclusion
Detection theory - methodsDetection theory - methods
Edgebased Gradient operators Linking Edge
Thresholding Constant, Addaptive,
Stastistical Relaxation
If this is a fish pixel, then…
13
Cross Filter Detector (CFD) Cross Filter Detector (CFD) aFilter 1
Variancec
Comparator
Filter 2 bEvaluator Traces
Signal a
Signal b
Signal c
Combine
Evaluator
Filter direction
CFD –Addaptive thresholdingCFD –Addaptive thresholdingMain challenge: Find the optimal threshold
signal
threshold
Detection methodsDetection methods
16
Foreground filter
Background filter
Comparator
variance
Evaluator
Background
Modelling
Comparator EvaluatorVideo
Echogram
Crossfilter detector
Common video processing
How to fit the Crossfilter to video like data?Can we learn something from the video world?
Background modelling. Background modelling. – the most important part. – the most important part.
Recursive Approximated median
Kalmann filter
Mixture of Gausians
Non recursive Previous picture Median Linear predictive Nonparametric
Background
Modelling
Comparator EvaluatorVideo
Common video processing
Background modelling. Background modelling. – the most important part. – the most important part.
Three best
1 Mixture of Gausians
2 Median
3 Approximated median
18
Ching , Cheung and Kamath found
Not much difference App. Median much faster
and simpler than the others
Sen-Ching S. Cheung and Chandrika Kamath Center for Applied Scientic Computing Lawrence Livermore National Laboratory, Livermore, CA 94550
EvaluatorEvaluator
Morfological filter Recognise fish on size and shape May use higher order statistics Connect parts of targets
20
Background
Modelling
Comparator EvaluatorVideo
Common video processing
21
CFD AND CFD AND DIDSONDIDSON
Tracking
3D approach
Detection methods
Echogram approach
Introduction
Summary
Inc.Video methods
22
Echogram approachEchogram approach
AmplitudeDetector
Gain96-Ch
Multi beam-viewer
Amp-Echogram
Multi 1 beamEchogram generator
23
Generate echograms and apply the Generate echograms and apply the Cross-FilterCross-Filter
a) Mean echogram At each range bin extract mean values from a selected number
of beams. Like an ordinary transducer with controllable opening angle
b) Max Intensity At each range bin, select the sample from the beam with
highest intensity
How to combine many beams into one ?
24
Generating Echograms from multi beamGenerating Echograms from multi beam
Data recorded by Debby Burwen
a) Averaging a number of beams 10x12 deg b) Pick the beam with strongest intensity
Many beams 1 beam
25
Testing the CFD on many to 1 Testing the CFD on many to 1 beam echogramsbeam echograms
Echogram approach
26
Echogram approach works well Echogram approach works well until density becomes too highuntil density becomes too high
We want to push the density limit
Echogram approach
27
CFD AND CFD AND DIDSONDIDSON
Tracking
Echogram approach3D approach
The original Cross filter
Aim, material and methods
Introduction
Summary
28
Adding a third dimensionAdding a third dimension Work directly on the multi beam data
Want to detect more than one target in the same range bin
3d-trace2d-trace
time
time
width
rangerange
3D approach
29
We added the beam dimension to We added the beam dimension to the filters the filters
DDF
New
Running window operators
2D 3D
Beam. nr
Range
Ping
Ping
Range
3D approach
32
Testing cross filter on a small Testing cross filter on a small trout in Fisha Rivertrout in Fisha River
Max Intensity echogram
CFD with filters CFD with filters and thresholdand threshold
Forefilt 3 x 3 x 3Back filt 3 x 3 x 3
Threshold Offset=20
36
CFD AND CFD AND DIDSONDIDSON
Tracking
3D approach
Detection methods
Echogram approach
Introduction
Summary
Inc.Video methods
37
Extended the background filter Extended the background filter with an approximated median with an approximated median operatoroperator
(N. McFarlane and C. Schoeld 1995)
ddfQ
1
1
BRBRFBR
BRBRFBR
AMthenAMQIf
AMthenAMQIf
38
And extended the And extended the comparator with comparator with
alternatives alternatives
Background
Foreground If ( a - b )>T )a
b
detectionThreshold
Background Background subtractionsubtraction
Forefilt 3 x 3 x 3Back filt 3 x 3 x 3
App.MedianThreshold Offset=20
40
CFD AND CFD AND DIDSONDIDSON
Tracking
3D approach
Detection methods
Echogram approach
Introduction
Summary
Inc.Video methods
41
The initial idea was to detect The initial idea was to detect traces directly by clustering traces directly by clustering
Cluster of overlapping fish pictures
( Work well in the echogram approach )
42
But data often showed traces But data often showed traces split up in individual fish picturessplit up in individual fish pictures
Clustering worked for big slow fish
Tracker needed for fast fish
Center of gravity
track
43
Special predictor can be made Special predictor can be made for multi beam datafor multi beam data
Special predictor can be formed from the DIDSON fish picture
In addition to traditional predictors are available such as Alpha Beta and Kalman
Fish center line
predictor
44
CFD AND CFD AND DIDSONDIDSON
Tracking
3D approach
Detection methods
Echogram approach
Introduction
Summary
Inc.Video methods
SummarySummary
45
Background
Modelling
Comparator EvaluatorVideo
Common video processing
Foreground filter
Background filter
Comparator
variance
EvaluatorEchogram
Crossfilter detector
DIDSON
Best method
Tracker3D-Foreground filter Comparator Evaluator
Background
Modelling
SummarySummary
46
DIDSON
Best method for moving targets
Tracker3D-Foreground filter
Background
Modelling
Comparator Evaluator
Needed in most cases Need for various predictors
depending on data
Improved foreground signal
Approximated Median
( a - b )>T )
a
b
3D better than 2DOptimise on improving foreground