|
Results
Test images
In the first project stage, a set of test images has been prepared by
using OLYMPUS CAMEDIA C-2040ZOOM digital camera. The image database contains
over 500 images of the rear views of various vehicles (cars, trucks, busses),
taken under various lighting conditions (sunny, cloudy, rainy, twilight, night
light). The image database is available as a .zip archive.
Preprocessing
Several preprocessing algorithms have been tested in early stage of the project,
by using Khoros Pro 2001 environment. The preliminary results, algorithm details
and instructions on using Khoros Pro 2001 have been presented in
documentation (available only in
croatian).
Enhancing the contrast in greyscale images
The test images have been converted to grayscale and a number of greyscale
image enhancement algorithms has been implemented and tested in order to obtain
better image segmentation - license plate detection and extraction.
In the first project stage the following contrast enhancement algorithms
have been implemented: histogram equalization, local histogram equalization
(CLHE - Constrained Local Histogram Equalization), frequency domain algorithms
and algorithm JGACE (Just Noticable Difference Guided Adaptive Contrast Enhancement).
The testing of algorithms indicate that the most efficient algorithm is
TGHE (Thresholding Histogram Equalization) which runs in 0.018 sec. on 640 x 480 images.
The most time-consuming algorithm was the three-channel filtering with non-linear
contrast stretching (0.555 sec. on 640 x 480 images).
The results have been evaluated by using some subjective as well as some objective
criteria. First evaluation was based on viewers' subjective judgment of license plate
readibility. The objective evaluation was based on the success of the subsequent
license plate detection algorithm. Based on the evaluation results, a strategy based
on gray-level histogram has been developed. For dark images the system uses
TGHE algorithm with threshold set to 10. If the gray-level distribution is
classified as convenient, the system uses the original grayscale image. In other
cases the unsharp masking with three-channel filtering is used.
A detailed description of the implemented algorithms as well as some
generated results can be found in
documentation (available only in croatian).
Some of the results are presented as PowerPoint slides (only in croatian)
(short version,
extended version).
A demo version of the program can be downloaded
(here). The implemented algorithms
can be tested on various images and with various parameter settings.
Short instalation notes are (here).
Active-vision based extraction of vehicle image
As a part of the third project stage a system for extraction of the vehicle
image from the scene has been developed.
The task of this system is the surveillance of some passage where vehicles have to
stop for some kind of control (customs border traffic control, parking lots - security
and access control, toll collection control etc.) and the extraction of the image
of the stopped vehicle. It is advantageous to capture the image containig only the vehicle,
with as high resolution as possible, in order to improve subsequent processing.
Several subtasks can be identified:
- vehicle detection;
- vehicle position estimation;
- image capture.
Such a system could be implemented in various ways. One approach is by using a
sensor (e.g. photoelectric or step-on sensor) for the vehicle detection and triggering
a camera to take a picture of the vehicle. As a triggering sensor we can use a camera
connected to a computer, and apply computer vision methods to detect the vehicle, estimate
its position and take a picture of it. In this implementation we use the latter approach,
as it enables better flexibility and accuracy. Another advantage is improving the subsequent
analysis because it enables taking the picture only of the most interesting part of the scene
- the vehicle alone.
The system performance is illustrated by an image sequence.
The program code is not available on the web because it depends on specific equipment
(pan/tilt gimbal for camera)
installed at the Department of Electronics, Microelectronics, Computer and Intelligent
Systems at the Faculty of Electrical Engineering and Computing in Zagreb.
A detaliled description of the implemented system can be found in
documentation (available only in croatian),
Recognition of alphanumeric characters in license plates
During the project phase IV, a system for alphanumeric character recognition
for license plate reading has been implemented.
The system is based on multi-layer perceptron, which takes individual
characters cut out from the thresholded image of the license plate to be read.
The system incorporates syntax analysis module which checks the recognized characters
according to the syntax rules used for license plates in various countries.
This module is able to correct some misrecognized characters, especially
some very similar and otherwise indistinguishable characters (e.g. "0" i "O" ili "1" i "I").
The system has been tested on our image database with correct recognition
ratio of about 90%. The recognition ratio could be improved by using finer
image resolution, particularly by directing the camera and zooming
only the license plate area. A detailed description of the implemented
system can be found in documentation
(available only in croatian).
Some of the results are presented as PowerPoint slides (only in croatian)
(short version,
extended version).
A demo version of the program can be downloaded
(demo).
Short instalation notes are
(here).
The compound system: automated license plate location, character extraction and
character recognition
A prototype system has been implemented by linking afore mentioned subsystems.
The system takes a digitized image of a vehicle, locates the license plate,
extracts the characters and recognizes them by using multilayer perceptron.
A demo version of the program can be downloaded
(demo).
Short instalation notes are
(here).
Publications
- Vlasta Srebrić, "Postupci poboljšanja kontrasta sivih slika",
Diplomski rad br. 1397 (mentor: prof. dr. sc. Slobodan Ribarić)
Fakultet elektrotehnike i računarstva, Sveučilište u Zagrebu,
Zagreb, rujan 2003., 87 stranica, 6 stranica dodatka
(PDF, in croatian).
- Goran Adrinek, "Segmentacija slike na temelju pokreta",
Diplomski rad br. 1392 (mentor: prof. dr. sc. Slobodan Ribarić)
Fakultet elektrotehnike i računarstva, Sveučilište u Zagrebu,
Zagreb, srpanj 2003., 123 stranice
(
PDF, in croatian)
- Kristijan Kraupner, "Uporaba višeslojnog perceptrona za raspoznavanje
brojčano-slovčanih znakova na registarskim tablicama",
Diplomski rad br. 1396 (mentor: prof. dr. sc. Slobodan Ribarić)
Fakultet elektrotehnike i računarstva, Sveučilište u Zagrebu,
Zagreb, rujan 2003., 104 stranice,
(PDF, in croatian)
- Josip Haluška, "Razvoj programskih sustava u okruženju Khoros Pro 2001",
Diplomski rad 1402, (mentor: prof. dr. sc. Slobodan Ribarić)
Fakultet elektrotehnike i računarstva, Sveučilište u Zagrebu,
Zagreb, veljača 2003., 93 stranice,
(PDF, in croatian)
- S. Ribarić, G. Adrinek, and S. Šegvić,
"Real-Time Active Visual Tracking System",
submitted to MELECON 2004,
(PDF)
|
