USA, and Australia, have applied computerized systems to simplify some tasks in ... License Plate Recognition (LPR) or Automated Number Plate. Recognition ...
Proceedings of the International Conference on Electrical Engineering and Informatics Institut Teknologi Bandung, Indonesia June 17-19, 2007
B-87
License Plate Recognition System for Indonesian Vehicles 1
1
Felix Arya , Iping Supriana Suwardi 1
Bandung Institute of Technology, Ganesa rd. 10 Bandung 40132, Indonesia
Nowadays, there is very few usage of information technology in Indonesian traffic system. Even among those select few, only a handful of the software were made by local developers. This is very unfortunate, especially since traffic is one of the fields that can reap a huge benefit from using information technology. In this paper, we propose the research and development of a software system to be used in Indonesian traffic system, in the form of a License Plate Recognition System In application, the system can be used to increase efficiency of several traffic-related services. Several applicable uses includes automated parking, traffic light surveillance, electronic toll collection, and vehicle surveillance for police uses. All of the mentioned uses have been proven to increase efficiency of service and reducing operational cost in the long run. The main problems in this research are the methods of acquiring a license plate data from an image. The problem of acquisition revolves around the accuracy, speed, robustness, and flexibility of the system. In solving the main problems mentioned above, several approaches are suggested. The main ideas are to define a fast method of locating plates in an image, and designing a robust and flexible algorithm of character recognition fit for Indonesian number plates. The expected result of this research is a system that is able to recognize license numbers accurately with acceptable speed and robustness.
1. Background The use of information technology in traffic management is something that has been discussed frequently in developed countries. Several countries, such as Canada, United Kingdom, USA, and Australia, have applied computerized systems to simplify some tasks in traffic management. In application, the use of computers in traffic management usually revolves around the use of artificial vision to substitute human eyes in several tasks. This substitution usually involves the recognition of individual vehicles by reading their license plates, and then using this information in other tasks. The reading of license plates in this automated way is called License Plate Recognition (LPR) or Automated Number Plate Recognition (ANPR). An example of traffic management using LPR is the “407 ETR” highway in Ontario, Canada. This modernized highway uses a computer system connected with cameras and transponders stationed on the highway’s entrance to list the vehicles entering and exiting the highway. The list is then used to charge the individual vehicles directly from their owners’ account. The system eliminates the need of having toll booths in every highway entrance, thus reducing the traffic jams caused by a queue on the toll-booths.
stationed with several CCTV (Closed Circuit Television) cameras connected to a computerized system. The system keeps track of vehicles in every intersection, records the traffic infractions, and in several cases, helped the police to locate a wanted vehicle. Research on LPR systems have been conducted in numerous institutions, commercial or educational, with good results. For example, one of such systems, developed by Liu Jilin and partners from the University of Zhejian in China is said to reach 97% accuracy of recognition [LIU01]. Even in nighttime, and with the vehicle in question moving as fast as 80 km/h, the system is said to maintain 90% accuracy, with relatively short processing time of 0.1 to 1.5 seconds per vehicle. It is unfortunate that the use of LPR system in computerized traffic management have not taken place in Indonesia. In an effort to begin using computerized traffic management, this paper will discuss the development of an LPR system specifically engineered to recognize Indonesian license plates. Future development of the system is expected to find usage in Indonesian traffic management system, and hopefully increasing the efficiency of these public services.
2. Analysis 2.1 Preprocessing
Another example is the traffic management system in London, UK. Almost every intersection in the city proper was
ISBN 978-979-16338-0-2
The first step to recognize license plates inside the image is the preprocessing. Firstly, the system would smooth the image by using a convolution matrix, and then apply a 5x5 median
657
Proceedings of the International Conference on Electrical Engineering and Informatics Institut Teknologi Bandung, Indonesia June 17-19, 2007
B-87
filter as suggested in [FED97]. These steps are taken to minimize the noise in the image that would otherwise hamper the localization process.
2.2 Plate Localization The inherent characteristic of Indonesian number plates is that the numbers on the plate differ by a sharp contrast with the background color. This is especially true for the most common license plates, which is comprised of white numbers on a black background. In a binarised image, this sharp contrast is easily noticeable through a horizontal projection of the image. The horizontal projection of an image with license plates in the m shows sharp spikes on the lines containing a license plate. These sharp spikes provide the system with lines in which it could search for plate candidates. The system scans the original image along these candidate lines to find a plate signature. Plate signatures are patterns in the brightness map of an image that represents a plate location [KWA02]. The pattern is formed by the relatively even spacing of characters in the license plate. Basically, the system finds the plate by applying a threshold to the original image, then examining the binarised image for sharp spikes in the horizontal projection. Along these spikes, the system searches for plate signatures and defines the plate candidates. After the candidates are defined, they are eliminated one by one by using the estimated plate size compared to the image. The elimination continues until there is only one candidate left. This approach has an obvious weakness in the fact that it can only recognize one plate in a picture. The presence of a second plate would result in one of them being eliminated by the elimination step, and thus, not recognized as a license plate.
Fig. 2. Plate Candidates The images above show an example of the localization algorithm used in the system. The first image is the original input image, and the second image is the original image after being analyzed by the plate localization algorithm. The blue regions on the second image are plate candidates, and the yellowed box is the last plate candidate left after elimination.
2.3 Character Segmentation After locating the plate, the next step to be taken by the system is to separate the characters in the plate to ease the character recognition process. Character segmentation is done by analyzing the horizontal and vertical projections of the localized plate [ZHA03]. In general, a noticeable spike or plunge in the vertical projection usually represents a gap between characters. Using these gaps as guidelines, the system separates the localized plate to vertical blocks.
Fig. 1. Sample input image
Fig. 3. Character Segmentation
ISBN 978-979-16338-0-2
658
Proceedings of the International Conference on Electrical Engineering and Informatics Institut Teknologi Bandung, Indonesia June 17-19, 2007
The above image shows the character segmentation algorithm at work. The system divides the extracted plate into vertical blocks (noted by red lines in the image), and prepares them for character recognition.
2.4 Recognition The last step is to recognize the characters inside the vertical blocks individually and arrange them into a valid license plate number. The system uses a multilayer feed-forward network with backpropagation to recognize characters. Vertical blocks produced by the segmentation routine are resized into images of 20x20 pixels large. The brightness maps of these images are then converted into input values for the network. The network has 400 input units containing brightness values of the vertical blocks, 50 hidden units, and 36 output units representing the 26 alphabets and numbers from 0 to 9. The training of this artificial neural network is also handled by the system. It receives the training set of specially formatted images containing characters with the approximate font used by Indonesian license plates. The default training procedure of this network is to limit its training to 10000 epochs. With a Pentium IV 1.6 GHz CPU, the training took around 10 minutes to finish. This relatively long training time is compensated by the system’s ability to save a trained network into an external file and loads it for the recognition process. This means the system only needs to do the training routine once, and when the trained network has been saved into an external file, it doesn’t have to do the training. The system has given up to 80% accuracy on several test cases, but the result is very deeply affected by the training set used. In some cases, with an identical image and a different training set, the system gave a very different result. To the point of this paper’s release, a universally acceptable training set has yet to be found.
B-87
concerns would be the automated calculation of the input variables to be used in plate localization and segmentation methods. When these methods are able to function without user intervention, the system should be able to operate almost fully automated. Another shortcoming to be addressed in future researches would be the recognition methods. Further researches in this aspect would be focused in finding a universally acceptable training set for Indonesian license plates.
5. Reference (1) Brown, Eric. (1992). Applying Neural Networks to Character Recognition. Northeastern University Internal Paper. 1992. (2) Clark, Paul and Majid Mirmehdi. (2000). Finding Text Regions using Localized Measures. International Conference on Pattern Recognition ICPR00, 2000. (3) Draghici, Sorin. (1997). A Neural Network Based Artificial Vision System for Licence Plate Recognition. International Journal on Neural Systems Vol. 8. 1997. (4) Federl, Pavol and J.R Parker. (1997). An Approach to License Plate Recognition. Proceedings of Visual Interface 1997. 1997. (5) Kwasnicka, Halina and Bartosz Wawrzyniak. (2002). License Plate Localization and Recognition in Camera Pictures. Artificial Intelligence Methods, 2002. (6) Jilin, Liu, et al. (2001). A High Performance License Plate Recognition System Based on The Web Technique. IEEE Transactions on Intelligent Transportation Systems, 2001. (7) Salameh, Walid and Mohammed Otair. (2006). Online Handwritten Character Recognition Using an Optical Backpropagation Neural Network. Issues in Informing Science and Information Technology. 2006. (8) Thillou, Celine and Bernard Gosselin. (2005). Combination of Binarization and Character Segmentation using Color Information, Proceedings of ICCVG 2005, 2005. (9) K.S, Yap, et al. (1999). Vehicle License Plate Recognition by Fuzzy ARTMAP Neural Network, Proceedings of WEC 1999, 1999. (10) Zhang, Yungang and Changshui Zhang. (2003). A New Character Segmentation Algorithm for License Plate. Proceedings of the 2003 Intelligent Vehicles Symposium, 2003.
3. Conclusion Though the system has given accurate results in several cases, it still hasn’t performed up to the expectations, especially if it were to be used in real time traffic management. The major downside to the system is that it still requires a lot of human supervision and input, especially on the plate localization methods. Another downside is that the result still depends heavily on the training set used. Despite those shortcomings, the system is a good basic for the development of a more advanced LPR system for Indonesian license plates. The localization and segmentation methods in particular are proven to be accurate and quite robust. With further improvements, the system should be able to operate in real time situations.
4. Future Developments Further research in the future would be focused on developing a more automated version of the system. The major
ISBN 978-979-16338-0-2
659