Transcript Presentation_Gural

MeteorScan Overview
and other
Transient Detection Algorithms
Pete Gural
[email protected]
Meteor Orbit Determination Workshop #3
April 17, 2010
1
Algorithmic Development Considerations
Imaging Modalities and Purpose
All sky – Fireball survey and meteorite recovery
Moderate FOV – Meteor flux, mass index, stream characterization
Telescopic – Ablation, orbits, spectroscopy, lunar impacts
2
Throughput
- Real-time, Near-real-time, or Post-collection
Detection
- Fast (high SNR) or robust (low SNR) algorithm
False alarms
- Tolerance for and mitigation approach
Computing
- Processing capacity, storage, interfaces
Analysis
- Calibration, Cueing and/or Science exploitation
Detection Algorithm Choices
Streak Detection
Matched Filter – Hypothesize motion, shift and stack, then threshold
Best Pd, Pfa but large hypothesis count limits the application to meteors
Hough Transform – Threshold pixels, transform to Hough space, find peaks  feed MF
Good Pd, Pfa suitable for near real-time with short latency
Orientation Kernel – Convolve spatial kernel, merge detections via temporal propagation
Good Pd, Pfa suitable for near real-time with short latency
Cluster Tracking – Threshold pixels, locate clusters, motion consistency
Moderate Pd, Pfa suitable for real-time tracking needing rapid response
Spatial Change – Threshold pixels and match to spatial signature
Poor Pd, Pfa useful when the transient leaves no temporal response
Background Removal
Clutter Suppression – Use noise statistics to whiten the imagery
3
Mean or Median
– Good for stationary background, lower noise threshold
Difference Frames
– Good for slowly drifting background, fast processing
MeteorScan 3.20 Overview
• Primarily for Meteor Detection in Video
– Limited analysis capability since users wanted to “roll their own”
– Operates at full resolution and near the recorded rate
• Used by the North American Professional Meteor Community
– Univ. of W. Ontario, NASA/MSFC, SETI
– Originally Real-Time on a Mac circa 1997
– Migrated to Non-RT on a PC/Windows system ingesting AVIs
• MeteorScan Capabilities
– Masking and FOV Calibration
– Detection via Hough Transform & MLE
– User confirmation review and editing
– Radiant association and statistics
4
– Software library for detection-only processing in Windows and Linux
MeteorScan Detection Processing
Noise
Tracking
Filters
(in blue)
Primary
Image Space
Secondary
Hough Space
Tertiary
MLE Space
<MLE>
MLE
Detect
?
.
.
Max
Likelihood
Estimate
.
Frame
Differencing
Primary
Thresholding
Hough
Transform
Hough
Peaks
Track
Hypothesis
Streak Detection - Hough Transform
Map spatial coordinate exceedance pixels into Hough space
lines that pass through each point


y
x
– Pixel Pair HT - two points define
line thus one point in Hough space.
Localize pairs to reduce ops count.
– Phase Coded Disk HT – convolve
PCD kernel around each point to
obtain orientation
MeteorScan
Traditional HT
3 points on a line
6
PCD
– Traditional HT – hypothesis all
Line in Traditional HT
(butterfly self-noise)
Pixel pair HT
N2 ops
SPFN - LFI
Phase coded disk HT
N ops
Confirmation Mode Screen Shot
7
MTP Detector: Croatian Meteor Network
• Video Compression via “SkyPatrol”
• CONOPS
– Save one RGB bit mapped file for every N seconds of video
– For each pixel, keep the max value in time and associated frame#
– Extending to temporal mean and std dev (excluding max) for flat fielding
• Max Temporal Pixel (MTP) meteor detection software
•
8
Uses the MeteorScan detection modules, Post-processing by CMN
Maximum Pixel Value
Frame Number of Max
Reconstructed Video
CAMS at the SETI Institute
• All-sky coverage with high angular resolution
• CONOPS
– 5 DVRs monitors 20 CCD cameras for motion detection at 2 sites
– Records all cameras via FTP compression (Flat-field Temporal Pixel)
– Download only compressed video snippets containing detections
• MeteorScan processed on DVR archive
•
Post-processing for triangulation and orbits by SETI
DVR
4 channels
DVR
4 channels
DVR
4 channels
DVR
4 channels
DVR
9
4 channels
Archived
Detections
via
MeteorScan
MeteorScan for Telescopic Meteors
• Fragmentation studies, Precise radiant positions
• CONOPS / Issues
– Very narrow FOV and large optics  deep stellar lm without intensifier !
– Meteor trailing losses still limits meteor lm  +6.5
– Small FOV lowers # meteors collected
– Orion 80mm f/5 finder scope
•
2x Focal reducer  2 degree FOV and stellar lm=+10.5
• MeteorScan has option for long streaks
5 km
10
Short Baseline Meteor Triangulation
Scott Degenhardt’s
“Mighty Mini”
Orion 50 mm
Transient Video Detection Applications
• LFI Detector for the Spanish Fireball Network
• Massive Compact Halo Object Detection
• Lunar Meteoroid Impact Flash Detection
• Meteor Tracking System
• Meteor Simulation for ZHR
11
LFI Detector: Spanish Meteor Network
• Large format CCD: 4K x 4K pixels
– All sky coverage with 2.4 arc-minute resolution
– Non-video system: stellar lm = +10, meteor lm = +2
• CONOPS
– Slow read out CCD  1 snapshot every 90 seconds
• Long Frame Integration (LFI) meteor detection
– Differenced frames ( stars + and -, meteors + or - ), Hough Transform PCD
– Post processing orbital reductions analysis by SPFN
-
12
=
HT
Massive Compact Halo Object Detection
• Jupiter sized objects wandering the galaxy
– Stars briefly wink out from occultation
– Find TNOs in the plane of the solar system
• CONOPS
– Collect pairs of dense star field video
– Search for short timescale occultation
– Use pair coincidence to rule out scintillation
• 2 Telescopes with frame rate CCDs
– Observation of an open cluster with good timing
• MachoScan to identify occulted stars
– Space-time coincidence of recorded AVIs
– Post processing analysis by Mount Allison University
13
Few meters
LunarScan: Lunar Impact Flash Detection
• Boulder Sized Meteoroids Smashing into the Moon !
– Hypervelocity impact creates a momentary flash
– Duration typically a few tens of milliseconds
– One lasted ½ second !
• CONOPS
– Monitor the dark face of the Moon
– 3 days around first and last quarter
– Minimum of two sites >20 km separation
• LunarScan software to locate flashes
– Register, Track mean and standard deviation
– Threshold, Spatial cluster
– Post-collection analysis by NASA/MSFC
14
AIMIT Meteor Tracking System
• Increase #s of meteors observed in narrow FOV instruments
– Enables spectroscopy and high resolution triangulation/orbits
• CONOPS
– Wide field camera cues steering system for narrow field instrument
• MeteorCue Detection Algorithm
– Threshold, Fast clustering, Centroid, Track, Mirror Commands
– Response time <100 msec (Galvo), <500 msec (Stepper)
–
15
Post-processing Univ of W. Ontario
MeteorSim Processing
Radiant
Particles assumed to have:
Initial direction along radiant vector
Random start position in cylinder
Fixed begin and end heights
Fixed magnitude
Initial speed V∞
Fixed population index r
Mag distribution = [-12,+6.5]
Undergone zenith attraction
Earth
Not decelerated
Distance fading loss
Atmospheric extinction loss
. ...
. .. .. . .
..
. .
Specific to CCD vs. Human:
Limiting magnitude
FOV geometry
FOV look direction
Resolution
Integration time
Angular velocity loss
Off-axis perception
Monte Carlo meteor influx simulation for video and visual observations/calibration
Converts video counts  Spatial flux  ZHR
Algorithmic Backup Charts
• MeteorCue
• LunarScan
• Streak Detection
– Matched Filter
– Orientation Kernel
– Fast Clustering
17
MeteorCue Processing
Mean, Threshold, & SNR
Tracking Filters
(Updated on a few rows per frame)
<X>
Full Frame
Imagery
30 fps
<X> + k1 
<SNR> + k2 SNR
Even Field
Odd Field
Row, Col, SNR
Row, Col, SNR
Alpha-Beta
Tracker
30 Hz
2 x 16-bit
Digital
Signals
Vx, Vy
Repeat
every
33 msec
Threshold
Each Frame
Cluster
Detection
Fast
Centroid
Tracker
Association
Update
Mirror
Commands
LunarScan Processing
Image
Courtesy
NASA/MSFC
Sept 16, 2006
Optional register (PCM translation),
Warp mean and  to current image
Threshold
Triplet + Doublet
cluster
detector
Exceedances
Update
Mean and
standard
deviation
Streak Detection – Matched Filter
Uses a “Track-before-Detect” approach
Remove Mean and Estimate 2nd Order Noise Statistics
Apply Covariance Inverse to Remove Clutter (Whitening)
Hypothesize Multiple Target Velocity Speeds and Directions
Shift Frames and Add for each hypothesis
Convolve with Smear Kernel
–
–
–
–
–
.
.
.
.
.
.
.. .
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
.
.
..
.
.
2
.
. ..
.
. . . . . 3.
.
.
. .
.
.
. .
. .
.
. .
. .
.
20
1
. .
..
. .
.
Mean Removal
Covariance Estimate
Clutter Removal
.
.
Velocity Hypothesis
Shift & Stack
Threshold Detect
Decluster / Culling
.
Multi-Frame Integration
. .
.
Streak Detection – Orientation Kernel
Small scale spatial-only convolution
– Convolve 8 orientation kernels across focal plane
– Detections are tested for temporal propagation
– Shown are 5x5 binary kernels (MetRec)



21
Can be higher fidelity with width and fractional fill
Can use larger dimensions  more kernels
Can be formulated as a spatial matched filter
Streak Detection – Pixel Clustering
Find Groups of Pixels (Limited Spatial Extent, Track in Time)
Row Indices
Threshold
Crossers
S
Column
Indices
1
0
1
1
3
1
3
0
1
1
2
0
1
1
2
5
1
1
0
1
4
8
1
3
Define Cell
Size from Max
Meteor Motion
Per Frame
Scale = 16 pixels / deg
Max = 51 deg / sec
30 frames / sec
Max  28 pixels / frame
Cell = 32x32 pixels
1
2
1
1
1
Remove Singletons - Fill 32x32 Cells with Threshold Crossers
22
Find Highest Peak Counts in 2 x 2 Cell Sums