license plate identification amir ali ahmadi jonathan neville justin sobota mehmet ucal
TRANSCRIPT
License Plate IdentificationLicense Plate Identification
Amir Ali AhmadiJonathan Neville
Justin SobotaMehmet Ucal
Outline
• Motivation
• Previous Work
• Approach
• Algorithms– Character Identification– Plate Extraction
• Results
• Conclusion/Future Work
Motivation
• Traffic Control
• Automated Ticketing
• Finding Stolen Cars
• High Speed Pursuit
Previous Work
• License Plate Identification/Recognition (LPI/R)– http://www.photocop.com/– Retrieves Plate Numbers for All States– Determines Speed– Several vendors
• Three algorithms for license number extraction
Previous Work
• Template Matching– Compares extracted characters to a set of
templates– Very reliable under standard conditions– Viewing angle, Lighting, plate size, etc. can
cause errors
Previous Work
• Structural Analysis– Uses geometric features and a decision tree to
determine character
– Very complex time-consuming analysis
Loops?# of Loops
Location of Loop?
Left Side Straight?B
8
yes
1
2 yes
no
no
top
bottom
middle
6
D
Previous Work
• Neural Networks– Trained by example– Adapt to characters’ distinctive feature– Performs well in bad conditions
Our Approach
• Template Matching
• Assumptions– Only white Maryland Plates– Camera angle directly behind car– 2 types of MD plates
• 6 characters with MD logo in center• 7 characters
Approach
Plate Extraction
Character Extraction
Template Matching
CharacterIdentification
Character Identification
Char. Extract
Support Set Extract
Comparison
Char.Filtering
TemplateFiltering
TemplateImages
LicensePlate
PlateNumber
Template Filtering
• Templates obtained from actual plates• Template Filtering
– RGB2Gray– Threshold (Black/White)– Resize
• Output array of templates
Character Extraction
• Plate resized to predetermined dimensions• Output array of extracted characters
Character Filtering
• RGB2Gray• Threshold (Black/White)• Median Filtering
Character Identification
Char. Extract
Support Set Extract
Comparison
Char.Filtering
TemplateFiltering
TemplateImages
LicensePlate
PlateNumber
Support Set Extraction
• Row sums• Column sums• Exclude low sums• Extract largest
continuous region• Resize to
template size
Comparison
?
?
Approach
Plate Extraction
Character Extraction
Template Matching
CharacterIdentification
Plate Extraction
• RGB2Gray• Threshold(Black/White)
• Row/Columnmeans
• Extract largestcontinuous whiteregion
Results for Character Identification
Input OutputLicense Identification
License Identification
License Identification
Results for Character Identification
Input OutputLicense Identification
License Identification
Results for Plate Extraction
Input OutputPlate Extraction
Results for Plate Extraction
Input OutputExtracted “M”
Failed Plate Extractions
Input OutputPlate Extraction
Failed Plate Extractions
Input OutputPlate Extraction
No Extracted Plate
No Extracted Plate
No Output
No Output
Conclusion
• Template matching approach was taken• Algorithm
– Plate Extraction– Character Identification
• Given the plates, we were able to identify almost all of the characters
• Plate extraction was limited to darker cars
Future Work
• Improve templates to better accommodate the plate characters
• Refine threshold levels for determining the whiteness in the picture
• Eliminate issues regarding glare, dirtiness of the plate, shadows, and white regions in the picture
• Dynamic character extraction – Character position found by the algorithm
Demonstration