Clarus Research: Visibility estimation from camera imagery

1
Clarus ResearchVisibility estimation from
camera imagery

• Robert G Hallowell
• Michael P Matthews
• 09 August 2006

2
Outline

• Why focus on cameras?
• Measuring Visibility
• Project scope / Methodology
• Results
• Summary / Future Ideas

3
Automating camera monitoring

• Massive investment in cameras
• ITS to Homeland Security
• High /time costs for monitoring
• Valuable road weather information
• Automated road weather algorithms
• Weather sensing
• Verification

State DOT Controlled Cameras
4
ASOS Visibility Sensor

• Pros
• Standard NWS reporting
• Estimates extinction coefficient
• Sensor site representative of 2 mile area
• Cons
• Tuned for aviation users
• 10 minute time filtering
• Approaching weather
• Limited density of ASOS stations

5
Camera Visibility

• Pros
• Camera directly on roadway
• Senses visibility similar to a human
• Impacted by local and approaching weather
• Next sensor within seeing distance
• Cons
• Only measures in one sector
• New research
• Q/C challenging

6
Camera Imagery for Visibility

• Catalog existing DOT camera sites
• Re-build MIT/LL camera processing algorithm
• Characterize sensor requirements
• Focus on key ranges of visibility (
• Extend to DOT camera imagery

7
Algorithm Flow Chart
Image Capture
Edge Detection
Image Registration
Composite Image Edge Generator
Visibility Algorithm
Normalized Edge Extraction Detector Calculations
Clear day composite image edges
8
Flight Facility Test Site
Roof (ff1)
Tiered deployment
COHU-3960 Camera
FF1 (30 meters)
Cab Level (ff2)

• Images started 2/10/06
• 33
• Mostly rain/fog

FF2 (5 meters)
9
Flight Facility Roof Camera View
Near Ridge 0.63 miles
Lincoln Lab 1.0 miles
Fourth Hangar 0.28 miles
Far Ridge 1.13 miles
Smoke Stacks 0.89 miles
Second Road 0.28 miles
Intersection 0.07 miles
10
Visibility Algorithm - Assessment
Estimators
ASOS Visibility (miles)
Mean Image
10
5-10
1-5
Edge Mean
0
0
110
80
Masked Edge
45
66
1-5
495
29
Edge Ratio
Video Algorithm Visibility (miles)
Edge Diff
5-10
4
193
105
119
Consensus
113
2
55
1698
10
Visibility Distance
MIT/LL Roof Camera 2/10 to 5/31/06 (estimators
trained on same data set)
11
Visibility Estimate during transition
12
Ground-level Vs. Elevated Camera
ASOS Visibility (miles)
10
5-10
1-5
1
0
1
2
149
1-5
75
335
92
Video Algorithm Visibility (miles)

• Overall results good
• Low vis events biased high
• Angle blurs edges
• Roof camera superior

5-10
4
387
19
32
7
7
14
646
10
FF2 road 3/16 to 5/31/06
13
Extending Method to DOT cameras

• Salt Lake City, UT - Archiving 6 cameras 1/10/06
  present
• Challenges
• Moveable cameras / Embedded text

14
Utah Cam48 Visibility Algorithm - Assessment
Building 0.35 miles
Overpass 0.78 miles
ASOS Visibility (miles)
10
5-10
1-5
Black Mt 9.62 miles
Ridge Line 11.52 miles
1
0
33
2
66
4
1-5
9
18
Video Algorithm Visibility (miles)
5-10
0
180
14
6
10
21
0
12
442
8,000 ft
Cam48 1/30 to 5/24/06 (808 images)
4,250 ft
Within 10 minutes, manual corrections made
15
Utah Cam56 Visibility Algorithm - Assessment
Small Hill 16.2 miles
Large Mt 17.2 miles
ASOS Visibility (miles)
Office Park 0.75 miles
10
5-10
1-5
Street Signs 0.17 miles
1
3
27
5
24
8
1-5
26
3
Video Algorithm Visibility (miles)
5-10
3
204
7
4
7,500 ft
10
21
9
13
436
4,250 ft
Cam56 1/27 to 6/1/06 (794 images)
Within 10 minutes, manual corrections made
16
Usefulness of unexpected edges
0.0092
0.0110
0.0075
0.0047
0.0019
17
Summary

• Truth - ASOS sensor has limits for tuning
• Visibility algorithm methodology verified
• High cameras better than low cameras
• Image meta-data needed
• Generic algorithm challenging

18
Future Work

• Continue algorithm refinement
• Work with high density camera network
• How to effectively use directional visibility
• Warning decisions
• Adjust for snow / rain drop signature
• Road condition detection

19
(No Transcript)