Efficiency of the Electronic License Plate Recognition Systems

The precise recognition of the vehicle determines the efficiency of intelligent transportation systems. Intelligent transportation systems become more effective if it can identify the vehicle. Electronic license plate recognition systems used to solve this problem. Proposed electronic license plate recognition systems model adapted for simple internet protocol video surveillance cameras. An analysis of the internet protocol video surveillance camera‘s processors, operating principle and software created guidelines for the algorithm. According to performance of the internet protocol video surveillance camera formed four efficiency criteria for license plate recognition algorithms. It's performed the experiment with four algorithms to search for a license number plate by different attributes of the license number. The most efficient algorithm was selected by criteria from experimental results. The algorithm is further analyzed and proposed the application model of the internet protocol video surveillance camera. DOI: http://dx.doi.org/10.5755/j01.eee.19.8.5396


I. INTRODUCTION
With constant growth of transport flows and increase in traffic jams, air pollution and noise level in the city increase, fuel is used without any need.In order to settle this problem the Intelligent Transportation Systems (ITS) are created and installed.The purpose of these systems is to collect information on traffic conditions and transport flows and transfer this information to transportation systems.ITS quality is mostly influenced by data collection and processing subsystems [1].ITS is considered more efficient, if such subsystems are capable of identifying a vehicle, i.e.ITS can trace vehicle, speed and direction of its movement and direct it to a faster route.Electronic Vehicle License Plate (VLP) recognition systems help to achieve this objective.The Electronic License Plate Recognition System (ELPRS) is a closed system, which is able to detect VLP from the received data.
At this moment the most effective measure of ITS for the data collection is video surveillance and processing systems [1].In such way ELPRS is implemented.In order to get lower system's resources expenditure and more accurate results in such ELPRS, the up-search of VLP is made in the image [2].At first a car is searched in the image.After finding vehicle, the VLP location search is performed.In the Manuscript received December 6, 2012; accepted May 24, 2013.location image VLP symbols are cut and the plate is read according to the optical recognition of symbols.The following is used for implementation of these tasks: intelligent video cameras and video digital signal processing (DSP).
ELPRS, implemented in the following methods, are sufficiently accurate and effective; however the disadvantages of such systems are high price and large energy consumption.The system uses a powerful processor, with integrated image processing and analysis functions.
This article proposes another way of realization of the ELPRS.A simple IP camera is monitoring traffic flow, capturing frames, where the most visible VLP.Of the frames found VLP and sent VLP picture of the network.
Such a realization will perform: analysis of system of IP video surveillance cameras; VLP location recognition analysis; application and research of VLP location recognition in the IP video surveillance camera.

II. ANALYSIS OF SYSTEM OF IP VIDEO SURVEILLANCE CAMERAS
IP video surveillance camera consists of: image sensor, video processing and codec subsystem, ethernet card, additional input and output interfaces.Video processing subsystem consists of video processing unit, read-only memory and random-access memory.

Efficiency of the Electronic License Plate Recognition Systems
Let's consider the video processing systems more.SoC (System on Chip) processors are used for IP video surveillance cameras (Fig. 1).Video processing subsystem performs video pre-processing (image scanning from image sensor, image conversion from RGB to YUV, image size changing, image split to blocks, etc.) and video postprocessing (image conversion to analogue signal, image precodec, etc.).Integrated co-processors are used for encoding.Peripherals consists of different processor inputs/outputs (USB, Ethernet interface, GIO, etc.), timers needed for systems, data transfer bus.The kernel of this processor is ARM processor, which surveys status of subsystems and changes their parameters.All subsystems communicate with processor via data bus [3]- [6].IP video surveillance camera with such processor codes the video, as shown in the diagram (Fig. 2).Depending on type of image sensor, processor scans the image in blocks or portions of rows.Video subsystems immediately convert the scanned video shot from RGB to YUV colour system; small video blocks are formed and stored in RAM.When the necessary number of shots is collected, encoding of the video is started: DCT coefficients and movement vectors of small video blocks are found.This data is encoded.ARM processor puts the encoded information to the network (RTSP, HTTP) packages, from which video data transfer flow is formed.Since the search of movement vectors is necessary during video encoding, movement recognition function is implemented according to movement vectors, which is performed by ARM processor [4].
Due to constantly improving video codec and processing algorithms, IP video surveillance camera software is constantly renewed.Due to this criterion and easier allocation and management of system resources, IP video cameras use Linux operating system, which has all interfaces with video processing processor's subsystems and periphery.A program is created in this system, which, according to system's status, sends instructions to ARM processor on how to control IP video surveillance camera.In order to improve or create new function for such type of IP video camera, we must create the program namely for Linux system [4], [5].
The main difference between DSP for video analysis (e.g.DM6467) [7] and video IP surveillance cameras processors (e.g.DM365) [6] is that one of kernels of DM6467 is video DSP, which is used for video analysis, and DM365 have only a minimum video subsystem, which is intended only for video codec.In other words, DM365 only ensure good quality video coding.
After analysis we formed the model of electronic ELPRS (Fig. 3) according to the analysed diagram of IP video surveillance camera (Fig. 2).As we can see, VLP location recognition will be performed by ARM processor.ELPRS will take video and information on moving objects from RAM.In order to implement such system, the following criteria must be taken into consideration: 1) Processor is not particularly fast and it uses 50-60% of its resources for video processing, therefore the method of VLP location recognition must be found with as few calculations as possible; 2) VLP recognition method can be performed only by ARM processor, because video processing subsystem only generates data for video codec, and coding processors only encode it; 3) VLP recognition method may scan the image and information on moving object from RAM; 4) Use as few resources of RAM for interim ELPRS results as possible.
Fig. 3. Proposed method of ELPRS on the video cameras.

III. VEHICLE LICENSE PLATE LOCATION RECOGNITION ANALYSIS
Human eye detects VLP number from the whole entirety according to VLP form and location on the car, its symbols and background colours, which brightness differs a lot.A person may read VLP from symbols only after processing of information.Image is stored in colour pixels in the video processing systems, and the algorithms of typical pixels search are used for recognition of VLP number.The same as human eye, VLP recognition methods search for pixels or entirety of pixels, which define VLP number's colour, outlines and colour of symbols, its form and place in the image.
Looking for the most efficient VLP recognition algorithm for IP video surveillance camera, we made the experiment with 129 photos with the size of 256×256 and vehicle in them with the visible license plates.During the experiment we observed reliability of algorithm i.e. how many numbers the algorithm has correctly recognized and evaluated it in percentage; we calculated the average duration of algorithm processing.We also calculated how many operations must be made for one pixel in order it could define VLP feature.In order to find out how much RAM the algorithm needs, during the experiment we observed the areas of pixels of data necessary in each step .

avg S
If in a specific step data is the entire picture, then we multiplied length and width.If only specific pixels are analysed, we add them all up.After that we calculated the average of all pixels of tests.
We carried out the experiment with the following algorithms: according to VLP colour [8]; according to VLP form [9]; according to outlines of VLP symbols [10]; according to corners of VLP symbols [11].
All algorithms were implemented with MATLAB standard functions and examined on computer with processor Intel Dual Core, 2×2.I.
In order to select the most effective algorithm of IP video surveillance camera, we evaluated each algorithm from 1 to 4 according to the analysed criteria.

S E
After such evaluation of all algorithms, we calculate its suitability for ELPRS system according to the following formula The calculated is shown in Table I, last row.As we can see, the highest is LP recognition according to corners; therefore we will apply it in IP video surveillance systems as ELPRS.

IV. APPLICATION OF VEHICLE LICENSE PLATE LOCATION RECOGNITION ACCORDING TO CORNERS IN ELECTRONIC LICENSE PLATE RECOGNITION SYSTEM
VLP location recognition according to corners consists of the following steps: search of outlines; search of corners according to minimum proper value; search of VLP possible places; VLP selection from possible places and its cutting from picture [11].
We will use 5 cycles for these four steps (2 cycles will be needed for the second step) in order to select the most typical pixels.During each cycle we will narrow the data processing interval and index the typical pixels.In such way we will lose RAM resources, but accelerate the process itself.We will reserve static matrixes in RAM, in which we will store indexes of typical pixels and interim results.
From the proposed ELPRS model (Fig. 3) we know that ELPRS system surveys the flow of video and fixes moving objects.If the object crosses the line in the road, the shot is fixed, which is immediately converted to YUV colour system and each component of pixel colour is saved in the memory in the matrixes , [ ] [ ] where n -video height, m -video width.Since video in YUV system is stored in 8 bits, therefore in ELPRS RAM we reserve I s bits

Movement recognition algorithm stores the area of moving object in the matrix
According to [12]- [17], Sobel edge detection algorithm is four times faster than the Canny edge detection algorithm.The edges detection algorithm is first filter to finding VLP, so this step will process most of the data.Therefore, we must choose the easier and the faster algorithm.Applying Sobel filter in the first step, we find all sides of in the area of video , Y M defined by .

O M
We apply Sobel filter both, horizontally and vertically, and store the results in matrixes , x G y G of the same size as Y M : where: 1 2 1 ( , ) 0 0 0 , As we can see from formulas 10 and 11, , x g y g may acquire negative values and negative result may exceed 128, therefore 8 bits are not enough for storage of values.However, knowing that in further acts these numbers will be squared, we straight multiply the negative values by -1 and allocate in RAM for these two matrixes G s bits We also apply the filter in the first cycle, which would pass only those pixels, which satisfy the following condition min ( , ) . x After indexing pixels, which define possible places of corners, the second step may be performed in these pixelsto find out proper values of the pixel matrix M and if its minimum proper value is higher than ,

C
T then the pixel is the corner.Therefore in the first step of the second step we will find the minimum proper values of the matrix .
M Pixel matrix M is defined where 2 , here G -Gaussian filter.We calculate the proper value min : We store the received result in the matrix .
λ M Since during calculation of proper value of the matrix Gaussian filter is used, which numbers are real, then for storage of the matrix λ M values 32 bits will be needed, therefore this matrix will need λ s bits In the second cycle of this step we will look for corners, which indexes of columns we will store in the matrix C M and the array .

C
A Using E M indexes we start the search in the matrix λ M with the scrolling window, which centre is the analysed pixel and the length and width is equal to odd number .

C R
This window scrolls from top to bottom starting from the left.If window value is higher than ,

C
T it is checked if there are higher values in the window environment than the central pixel.If this condition is satisfied, we consider that the central pixel is the corner.Consequently: .
16 bits are needed for storage of array C A value and only 1 or 0 are stored in the matrix , C M therefore bits will be needed for storage of corner indexes, which depend on the width of window 32 2 2( 16) .
After finding the corners and storing their indexes in the matrix bits.Using the indexes of corners, we start the cycle of the third step from the first found bottom corner.We consider this corner the left corner of VLP.Using the scrolling square window, which side is equal to ; 1 2 + = r R L its movement direction is shown in Fig. 4. At first it moves to direction 1 and looks for the corner.If the corner is found, 1 is added to C and the new corner becomes the window centre.After that the next corner is searches.If there are no new corners in the window, the window moves through 5 more pixels to the last movement direction 1.If the corner is still not found, it is fixed that the last found corner is the top left VLP corner and we start to move in the direction 2 and look for new corners.If window cannot find new corners also moving to direction 2, it is fixed that this is not VLP.In such way the bottom right corner of number is searched moving to direction 3.
The quantity of RAM needed for interim results is calculated according to the (25).

V. EXPERIMENT RESULTS
We were played during an experiment video (711x400 pixels) with traffic on a motorway.Motion detection algorithm cut moving vehicle in the image from the video frame.The ELPRS algorithm is tested on three systems: PCs with Intel Dual Core, 2 x 2.4 GHz on the DM365 board leopard and Freescale i.MX27 Evaluation other.Select these optimal parameters [13]: Of the 969 images with the vehicles the ELPRS algorithm correctly cut 941 numbers.The experiment was monitored in the intermediate results: area of the vehicle in the pixels, the vehicle's contour pixels area and the vehicle's corner's pixels area.The results are sorted from the lowest to the highest according to the moving vehicle area and drawing graphs (Fig. 5-Fig.7).From these results, we have chosen the maximum and calculated according to the memory used, which is shown in the Table II.In addition, this table includes reserved memory to operating memory, and calculated the ratio between used and reserved memory.We have arranged the obtained results according to the moving object area, and featured the graph (Fig. 8).
In addition, we have compared the same algorithm, when the license plate search is performed with the proposed indexing and without it (Fig. 9).

VI. CONCLUSIONS
The electronic license plate recognition system can be offered for intelligent transportation solutions, parking systems, as the clipped image of the license plate is about 20KB.Therefore, the character recognition we believe the central intelligent transportation system server.The proposed algorithm is suitable for image processors, as it is reliable (97.11%) and fast: search of the license plate took 12ms on a PC, DM365 -207 ms i.MX27 -165 ms.
A graph shows, that the processing time increases of algorithms exponentially, when increasing the area of the moving object.Compared with the time graphs of the algorithm with proposed area indexing and without it, the PC saves time for 1ms, so the image SoC will save approximately 8ms.The disadvantage of the proposed algorithm is a demanding operating memory(2.75MB), although they are actually using only 43%.
As we see from Table II, it is possible to reduce a size of A array save index of the searching area.The used corner search method of the minimum eigenvalue is not the best, because the eigenvalue need to save the most of the memory and complex calculation.Therefore, the proposed vehicle license plate search, according to the methodology of the corners, should be adapted to different corners detection algorithm.
4 GHz.The received results , we filled in the Table

ALM
we can start the search of VLP location.We have stored the possible locations in the matrix , which is made of 100 rows and 5 columns.The first element of the row is number of corners; the second and the third column store the coordinate of the left bottom VLP location corner, and the fourth and the first -the coordinate of the top right corner.This matrix will need 8000

1 ) 3 )M
Finally the window returns to the left bottom corner in the direction 4 and all corners C are counted.If , min C C > we fix that this is the location of VLP and save in the matrix the number of L M corners and the coordinate of the left bottom and the right top corners of VLP.We delete the corners found in case of successful or unsuccessful location search from the matrix, After founding all possible locations of VLP, we select from them the most suitable according to the following criteria:2) The actual VLP location must have one of the most corners; The actual VLP location must occupy the area, which length is as close as possible to w and the heightas close as possible to h ; 4) The actual LP location must be as close as possible to the video bottom.At first the possible locations are analysed according to the first 2 criteria.If several locations are found that meet the criteria, then they are evaluated according to the criterion 3. The found VLP location is cut from the matrixes , JPG image is formed from the data of these matrixes and is sent by Internet.

Fig. 5 .
Fig. 5. Filled pixels in the image of the object.

Fig. 6 .
Fig. 6.Filled pixels in the image of the corner.

Fig. 7 .
Fig. 7. Filled pixels in the image of the corner.
The experiment was recorded license plate search time: we begin to calculate a time, when loaded , memory and stop a time, when excluded the license plate.

Fig. 9 .
Fig. 9. Process duration of the VLP location search with (red, first from the top) and without (blue, second) area's indexing.

TABLE I .
COMPARISON OF SEARCH ALGORITHMS OF VLP LOCATION.

TABLE II .
MEMORY ALLOCATION.