Video-based License Plate Reader Qing Li (SCPD) Department of Electrical Engineering, Stanford University
Method
Motivation License plate reader, also known as Automatic number plate recognition (ANPR), was first invented in 1976 in at the Police Scientific Development Branch in the UK and now widely used in numerous real-life applications, such as traffic control, automatic toll collection etc. These system usually requires special hardware: 1 or more cameras with fixed viewing angle, field of view and so on. Many effort has been made to make ANPR more reliability and accurate nowadays. But it’s also gets interest beyond the police forces as personal camera and the computing hardware gets cheaper. My project is to detect and recognize the license plate by analyzing videos from personal devices, such as cell phones and camera carried by hobby drones, which usually has low resolution or speed limit.
Challenges Perspective/ viewing angle
color
illumination
frame
4 6 8 10 12 14 16 18 5
10
15
20
25
30
Low resolution
35
Motion blur
Read video
Hough transform
-30
Special character
“false” plate
Related Work and reference 1. Shan Du et al. “Automatic License Plate Recognition (ALPR): A State-of-the-Art Review”, IEEE Transactions on Circuits and Systems for Video Technology, Volume:23 , Issue 2, Page 311, 2012 2. K. V. Suresh et al., “Superresolution of license plates in real traffic videos,” IEEE Trans. Intell. Transp. Syst., vol. 8, no. 2, pp. 321–331, 2007. 3. Previous project in this class in 2009 did the license plate detection with different method (no post processing and letter/number recognition).
-20
original image
-10
20
cut noise • Segment letter /number portion
100 200 300 400 500
Plate detection •
600
800
1000
1400
1600
o o
1 Next ? Alarm? Multi-frame?
0
0
40
0
50
60
50
Plate image convolve with template 8 10 20 30 40 20
40
60
80
100
120
140
Max profile for template 8
1.2 1 0.8 0.6 0.4 0.2 0 -0.2
Experimental Results
• Results: License plates were detected successfully and highlighted. The sub-image is corrected and shown as intermediate results. Final result is print and displayed. • Evaluation: frames, angles, illumination, odd characters can be addressed, although results depends heavily on resolution.
-50
Manual made template Pre-run to test size Peak detection and rejection
o
1800
30
20
Recognition Template matching 1200
20
60
20
400
60
40
800
1000
40
20
40
200
• • •
20
700
Extract by Boundary/Edge Vertical edge1 Edge density Boundary ratio Extract by Color feature 2 Extract by Texture feature Extract by character features (SIFT)
10
60
60
900
0
40
600
Note, iphone was used to take all video samples.
2
Standard plate VS …
Sub-image • Hough Transform to de-skew correction • Edge detection to
License plate detection from video frame 540
0
50
100
150
License plate detection from video frame 1 (IMG_5885.MOV)
100 200
100 300 200 400
(IMG_2298.MOV)
300
1
500
2
400 600
• Future work: can reconstruct higher resolution signal from low resolution images by combining enhance signal and reliability by combining neighour frames 500
700
600
1
800
700
900
800
1000
900
200 400 600 Plate extraction
1000
400 600 Plate extraction
200
800
1000
1200
1400
1600
Autocorrection
1800
10 20 30
20
40
60
1000
1200
1400
1600
1800
Autocorrection
5 10 15 20 25
10 20 30 40 Plate extraction
Recognized as 3FEJ331
• Future work: can reconstruct higher resolution signal from low resolution images2 by combining neighbor frames to enhance signal and reliability.
800
Recognized as 6TPP216 Autocorrection
20 40 60 20 40 60 80 100 120 140
Recognized as 7FYJ988
