vehicle detection with satellite images presented by prem k. goel ncrst-f, the ohio state university...
TRANSCRIPT
Vehicle Detection withSatellite Images
Presented by
Prem K. Goel
NCRST-F, The Ohio State University
Workshop on
Satellite Based Traffic Measurement
Berlin, Germany
9-10 September 2002
Image Processing Algorithms: Performance Evaluation
Acknowledgment C. Merry, G. Sharma, F. Lu,
M. McCord,
Past students: P. Goel, and J. Gardar
Vehicle Identification in High Resolution Satellite Imagery
• Infrequent Image Acquisition from satellites
• Stereo Coverage May be Unavailable
IKONOS Satellite Imagery: Tucson, AZ
Zooming-in
Image Segment for Processing
Zoomed and Pan Satellite Imagery (Columbus)
Problem Statement
• 1-m resolution image• 8 or 11-bit data• To detect and count
vehicles• Vehicle classes – cars
and trucks• No road detection
Pavement Background Image
• Lack of stereo Images
• Background (Pavement) Image
• No Background
• Background Based•Bayesian Background Transformation (BBT)•Principal Components (PCA)
•Gradient Based
BBT Method: Flow Chart
Update probabilities
Highway Image (I) Background (B)
Background Transform
Estimate Distribution Parameters
Threshold
Clustering and other operations
Vehicle Counts
Converged?
Yes
No
•Estimate probability of a pixel being stationary based on change from background
Distributions of gray-levels in two classesInitial prior probabilities
Principal Components (PCA) Method
Principal Components Analysis
Binary Image
Vehicle counts
Roadway only Image (I) Background (B)
S = I + B
D = |I – B|
V1=Var2x2(S)
M1= Mean2x2(S) M2= Mean2x2(D)
V2=Var2x2(D)
Select PC Band. Threshold
Clustering and other operations
PC Bands 1-4
PCA-based Method•Bands to capture texture and change•Re-orient bands
Segmented Highway Image (I)
Calculate Gradient Image
Threshold
Morphological operations and Clustering
Vehicle counts
Gradient Based Method•The ‘edge’ at vehicle boundaries•Gradient image = image with two classes
Threshold-try to incorporate spatial distribution of gray values
Gradient based method
OriginalImage
Binary Image
Final Outcome
Simulated Images
• No Method was best• Different method performed well for different images• Performance Evaluation on Real Images crucial
• General Characteristics– Vehicles vs. pavement
• pavement type, vehicle color, atmospheric conditions
– Objects: Road signs, Lane markings– Road geometry– Traffic density
Real Image Test Cases
Image: I 75 – 1
Main Characteristic•Pavement material transition
Thresholded PC Band
Clustered Thresholded Gradient Img
Clustered
I 75 – 1
Probability Map Clustered
Probability Map
Image: I 75 – 2
•Pavement material transition
Thresholded PC Band
Clustered Thresholded Gradient
Img
Clustered
I 75 – 2
Probability Map
Clustered Probability
Map
Image: I 270 – 1•Pavement material transition•Overpass•Lane markings•Curved road segment
Thresholded PC Band Clustered
Thresholded Gradient Img
Clustered
I 270 – 1
Probability Map Clustered Probability Map
Image: I 270 – 2
Thresholded PC Band
Clustered
Thresholded Gradient Img
Clustered
•Lane markings•Pavement material transition•Straight segment•Fairly dense traffic
I 270 – 2
Probability Map
Clustered Probability Map
Image: I 70 – 1
Thresholded PC Band
Clustered
Thresholded Gradient Img
Clustered
•Lane markings•Sign board•Fairly dense traffic•Straight road segment
I 70 – 1
Probability Map
Clustered Probability Map
Image: I 10 – 1
•Straight road segment•Median•Good vehicle vs. pavement contrast
PC Band Thresholded… ClusteredGradient Img Thresholded… Clustered
I 10 – 1
Probability Map Clustered
Image: I 270 – 3
•Multiple pavement material transitions•Median•High traffic density
I 270 – 3
Image: I 71 – 1
•Poor vehicle vs. pavement contrast•Illumination change•Overpass
Thresholded PC Band
Clustered
I 71 – 1
Thresholded Gradient Img
ClusteredClustered Probability Map
I 71 – 1
Image: I 70 – 2
•Cloud cover•Overpass•Pavement material transition
I 70 – 2
Thresholded PC Band
Clustered
I 70 – 2
Thresholded Gradient Img
Clustered
I 70 – 2
Probability Map Clustered Probability Map
I 70 – 2
Results SummarySummary: Errors of Omission and Commission
•BBT and gradient method give numbers close to the real values•Large errors of omission and commission for PCA and gradient based method•Low omission and commission errors for BBT method
Summary
Future Needs
• Methods Not Requiring Background
• Post-processing
– sieving and clustering– Effort– Process