Sunteți pe pagina 1din 5

Proceedings of the Eighth International Conference on Machine Learning and Cybernetics, Baoding, 12-15 July 2009

INTELLIGENT CAR-SEARCHING SYSTEM FOR LARGE PARK


HUA-CHUN TAN, JIE ZHANG, XIN-CHEN YE, HUI-ZE LI, PEI ZHU, QING-HUA ZHAO Department of Transportation Engineering Beijing Institute of Technology, Beijing, P.R.C., 100081 E-MAIL: tanhc@bit.edu.cn

Abstract:
Find a car in large park is a challenge for intelligent parking lot management system. In this paper, an intelligent car-searching approach for large parking lot is presented. In the new approach, some cameras are set up in each road. Vision information of the car, including car color and license plate are recognized and saved in database. Considering that no license plate recognition system can 100% recognize all plates, an intelligent car searching algorithm based on the similarity of license plate recognition and color classification probabilities is proposed to sort the retrieval results. The new approach is tested in a large outdoor park. Preliminary experimental results show the efficiency of proposed car-searching system.

Keywords:
Intelligent car-searching system; License plate recognition; Color classification; Database management system

1.

Introduction

Parking lot management plays an essential role in ITS (intelligent transportation systems). A parking lot management system is only a charge system which can count the number of parked cars and their elapsed time by paper cards in the beginning. With the development of ITS, computer vision and other related techniques, an intelligent parking lot management system, would not simply charge fee, but also should be designed to process more and more functions, including intelligent parking guide [1,2], [8], car parking occupation information [3-5], [9] and reverse car-searching [6,7], etc. In these systems, many of them focus on enhancing searching function more efficiently for available parking spaces. Goals of enhancing searching function include counting the number of parked motor vehicles, monitoring the changes of the parked vehicles over the time, and identifying the location to guide drivers to find vacant parking spaces efficiently [1-5, 8-9]. However, it is difficult for a parker to find his car because of similar situations and more floors in a large park.

Some researches try to deal with this problem. Beijing UnisPark Technology Co. Ltd [6] developed an intelligent reverse car-searching system (IRSS) to search a car in large park. They report that two kinds of methods are used in IRSS: license number identification technique and locating by card reader technique. In the first method, the car is localized by license plate recognition system. The second method find a car by card reader terminals which are initialized over the park. Xiamen Baoneng Technology Co. Ltd [7] developed an intelligent car-searching system based on vision information. In the approach, some cameras are installed on each road to track the passed cars by license plate recognition system, and record their travel paths. Then the drivers can search for their cars by license plate. These methods are efficient if the license plates are recognized correctly, or drivers inform the card reader before he leaves the park. However, drivers always forget going to card reader to record their location. For license plate recognition system, no one can claim that it can 100% recognize license plate. In many cases, if the license plate recognition system fails to record the correct license plate, the car can not searched properly. In this paper, a new method for car searching system is proposed based on vision information. The system is similar to the method proposed in [7], and improves the performance in three points: Firstly, considering that no license plate recognition system can 100% recognize all license plates in all conditions, the information of license plate recognition and car color classification are both used for car searching to improve the robustness of car-searching system. Secondly, the similarity of vision information to the input conditions, including license plate and car color, are jointed to sort the retrieval results. Thirdly, considering the license plates may not be recognized correctly and result in great deal of invalid records, delete invalid records process is also designed to improve the retrieval speed. An intelligent car-searching system based on the proposed method is developed and tested in a large outdoor park. Preliminary experimental results show that the method can search cars correctly in the first 10 records. The rest of the paper is organized as follows: section 2

978-1-4244-3703-0/09/$25.00 2009 IEEE 3134

Proceedings of the Eighth International Conference on Machine Learning and Cybernetics, Baoding, 12-15 July 2009 describes the system architecture and steps including recognition process. Section 3 discusses the proposed algorithm for car searching. Section 4 shows the experimental results and finally section 5 concludes the paper. 2. System overview

The proposed system will be used as a part of parking lot management system. An overview of car-searching system for large park is shown in Fig. 1. The processes are briefly described in the following. 2.1. Image capture
Vehicle Detection

The cameras are set up in each road intersection to supervise the passed vehicle. When a car passes the camera, the camera detects the vehicle automatically, and triggers the image capture process. 2.2. Vehicle detection In this step, the vehicle is detected by motion detection and verified by Adaboost-based method. Since the cameras are installed in fix location, background images are generated firstly to detect motive object. Then Adaboost-based algorithm is used to verify if the motion object is a car. If the car is detected, location and size of the car are recorded for license plate recognition and color classification processes. 2.3. License plate recognition Based on the information of vehicle detection process, the license plate is detected by auto-correlation based method proposed in [10]. Then the characters are segmented using projection-based method proposed in [11]. Finally, all segmented characters are recognized by pattern-match based method, except that the first Chinese character is recognized using a BP network classifier. Considering no one can claim that the license plate recognition system can automatically recognize all license plates in 100% accuracy, we record all probabilities of character recognition results in database. These probabilities are important for the following car-searching process.

Car Searching

Car-Color Classification

License Plate Recognition

Figure.1

Framework of proposed car-searching system

2.4. Color classification Based on the location and size of the detected vehicle, the area of the car is extracted. Considering the car may have two or more colors, or the area contains large area of background since the vehicle is not extracted accurately or is affected by illumination, the colors are clustered firstly to extract the majority color in this area. Then the majority color is classified to one of 10 types of pre-defined color by a Bayesian classifier. These pre-defined colors and their pre-probabilities are shown in Table 1. The probabilities of color classification are also recorded in database for car-searching process since the color may be inaccurately recognized.
Table 1 Pre-defined color and their pre-probabilities of cars (%)

Color Pre-Probability Color Pre-Probability

Argent 30.9 Green 4.8

White 27.4 Yellow 3.5

Blue 9.8 Brown 3.4

Black 9.6 Gray 2.8

Red 5.5 purple 0.6

3135

Proceedings of the Eighth International Conference on Machine Learning and Cybernetics, Baoding, 12-15 July 2009 2.5. Car searching process and Data management When a parker forgets the location where his car is parked, he can ask for car-searching system by inputting his license plate number and/or color of his car. The system would show the retrieval results according to the confidence level between the input information and the saved information in the database. Considering the records are saved according to recognized license plates, if a car is wrongly recognized, the database can not delete the record even after the car leaves the park. Thus many invalid records would affect the searching speed. To improve the efficiency of database management, we propose to delete these records by cars parking time. The details of this process are described in the next section. 3. Proposed car-searching method match results since the recognized results may be incorrect, but also calculate the similarities among the input conditions and all the saved recognition probabilities according to Equ.(1), where S i means the similarity between input conditions with the ith record in database, p ij is the probability between jth character of the ith record with the input jth character, q i is the probability between the color of the ith record with the searching color.

S i = p ij * q i
j =1

(1)

The proposed car-searching approach is implemented after the processing of captured image, plate recognition and color classification. The proposed method is based on the three reasons: The proposed car-searching approach is implemented after the processing of captured image, plate recognition and color classification. The proposed method is based on the three reasons: The license plates are recognized in a probability. No one can recognize all plates in 100% accuracy since various conditions, such as illumination, occlusion, etc. may occur; Thus, searching a car only by matching license plate may bring about wrong results. The color is an important feature for a car. Though it is not a unique feature for searching a car, the color may improve the robustness of car-searching process. Usually, the license plate is used as identification of a record. Since the license plates may be wrongly recognized, using the recognized license plate as identification may result in the record can not delete when it is recognized correctly on exit, and the system think the car is still in the park. Furthermore, such dead records may increase and affect the efficiency of car-searching system after a period time. To improve the performance of car searching system, we propose a method to deal with above problems. In the new method, both license plate and color are used for car searching. When user inputs search condition, including license plate and color, the system not only provides the

For example, after character recognition, we can find that the last character of a cars license plate is 1 with probability 80%, 2 with probability 10%, etc., and its color is yellow with probability 30%, white 50%, etc. When a user searches a yellow car with last character 1, the system would read the record and find qi is 30%, p 7 is 80%. After read all probabilities by input conditions, the system calculates the similarity between the record and input condition. Finally, user would get the first 10 similar records sorted by S i . To improve the efficiency of database management, we propose to delete these records by cars parking time. If a record is not changed in a long time, the system will tell the park management to check this record. The management will delete it if the car has left park to improve the efficiency of car-searching system. 4. Experimental Results

To test the performance of proposed method, three videos in an outdoor park are captured. Fig. 2 shows an example in pre-process. The image with vehicle is captured as Fig. 2(a), the vehicle and the license plate are detected as Fig.2 (b). Based on the location of the car, the colors are clustered into 3 classes, as shown in Fig. 2(c). About 200 vehicles are detected in the experiment. Based on the result of pre-process, license plate and color of the car are recognized. All probabilities are calculated and saved in database. The proposed car-searching system has been prototyped using C++ and tested in an outdoor large park. Fig. 3 shows the interface of this system, where Fig.3(a) is welcome page, Fig.3(b) is the car-searching condition input interface and Fig.3(c) is the search results. When user input his search condition, the system will give the first 10 results sorted by the similarities between the record and input condition. In our preliminary experiment, all cars can be correctly retrieved in the first 10 results.

3136

Proceedings of the Eighth International Conference on Machine Learning and Cybernetics, Baoding, 12-15 July 2009

(a) (a)

(b) (b)

(c)
Figure. 2 Example of results in pre-process (a) Captured image (b) Results of vehicle detection (c) Results of color clustering

5.

Conclusion

(c)
Figure. 3 Interface of car-searching system (a) Welcome page (b) car-searching condition input interface (c) search results.

Spending much time for parkers to find their car in large park is quite popular in our daily lives. In this paper, we present an intelligent car-searching system based on vision information. The information of license plate and color are both used to improve the robustness of the system. The probabilities of recognized license plate and color of the car are considered to sort the retrieval results. The preliminary experimental results show the efficiency of proposed car searching method. Acknowledgements This work is supported by Beijing Training Programming Foundation for the Talents (20081D1600300343), EYSCR fund of BIT (2007Y0305)

References [1] Lee. Jung-Wook ,Yoo.Yoon-Bong , Rho. JaeJeung ,Choi. Sae-Sol, An enhanced parking lot service model using wireless sensor network, Proceedings - IEEE INDIN 2008: 6th IEEE International Conference on Industrial Informatics, pp. 349-354, July. 2008. [2] Sheng-Fuu Lin , Yung-Yao Chen , A vision-based parking lot management system., 2006 IEEE Conference on Systems, Man, and Cybernetics.Peter. C, pp. 2897-902, Oct. 2006.

3137

Proceedings of the Eighth International Conference on Machine Learning and Cybernetics, Baoding, 12-15 July 2009 [3] Caliskan. Murat , Barthels. Andreas , Scheuermann. Bjrn , Mauve. Martin, Predicting parking lot occupancy in vehicular ad hoc networks, 2007 IEEE 65th Vehicular Technology Conference, pp. 277-81, April 2007. [4] D.B.L. Bong, K.C. Ting, K.C. Lai, Integrated approach in the design of car park occupancy information system (COINS). IAENG International Journal of Computer Science, Vol. 37, No. 1, pp. 7-14, March. 2008. [5] Fabin, Tom. An algorithm for parking lot occupation detection. Proceedings - 7th Computer Information Systems and Industrial Management Applications, CISIM 2008, pp. 165-70, June 2008. [6] Beijing UnisPark Technology Co.,Ltd .Available: http://www.unisapex.com/Solution/MesRead.aspx?Ui d=22 [7] Xiamen Baoneng Technology Co.,Ltd .Available: http://www.goepe.com/apollo/prodetail-xmbnkj-24161 9.html [8] Leephakpreeda, Thananchai, Car-parking guidance with fuzzy knowledge-based decision making. Handb. Environ. Chem. Vol. 5 Water Pollut, Vol 42, No. 2, pp. 803-809, February. 2007. [9] Idris. M.Y.I. , Tamil. E.M. , Razak. Z. , Noor. N.M. , Km. L.W, Smart parking system using image processing techniques in wireless sensor network environment, Information Technology Journal, Vol 8, No. 2, pp. 114-127, 2009. [10] Hua-chun Tan and Hao chen, A novel car plate verification with Adaptive Binarization method, ICMLC, 2008, 4034-4039. [11] Yinpeng Chen, Xiaoqing Ding, License plate location and character segmentation in comples vehicle images, Infrared and Laser Engineering, Vol.33 No.1, 2004, pp.29-33.

3138

S-ar putea să vă placă și