動態車牌辨識系統之研究

A Study of Motion Based Vehicle License Plates Recognition System

莊祥琳(Hsiang-Lin Chuang)

指導教授 ： 謝景棠

淡江大學/工學院/電機工程學系碩士班/碩士(2007年)

https://doi.org/10.6846/TKU.2007.00093

摘要

車牌辨識系統已廣泛應用於收費停車場、社區大樓停車場以及高速公路收費站，等車輛管理與監控系統上之應用。本系統是將攝影機架設在移動載具上持續拍攝前方的動態車輛並且偵測，此研究希望能夠與計程車聯盟合作，協助警調單位於贓車查緝或過濾可疑車輛以及協助銀行於權力車的搜尋。
已提出的系統中，一般是以固定架設攝影機的方式來擷取過往車輛的影像，且背景環境較為單純，本系統的不同處為動態偵測技術，其所偵測的街道背景較為複雜且經常改變，我們針對串流影像提供低運算量的演算法做分析，包括車尾偵測、影像清晰度判別、車牌位置偵測等，並且結合以上方法自動從輸入的影像資訊中擷取所需的車輛影像。同時利用串流影像連續的特性，可以對同一台車作多次的車牌辨識以增加成功的機率，比起過去靜態拍攝的方式更能實際應用於真實世界。
在本篇論文中我們運用離散小波轉換和類神經網路來完成汽車牌照辨識的應用。而本文主要分為幾個部份，第一個部份是車尾位置偵測，第二個部份是汽車車牌的定位，第三個部份是牌照中字元的切割，第四個部份是牌照字元的辦識。
本文所提出的方法中，我們使用Haar小波轉換將輸入影像分為LL、LH、HL、HH四個子頻帶，利用每個頻帶不同的特性來實現，車尾偵測、清晰度判別、車牌定位，可以去除雜訊、減少計算時間、和記憶體的使用量。此演算法有利於實現一個高效能的車牌定位系統。
車牌字元辨識部份，我們利用類神經網路的方法來辨識字元。類神經網路具有高容錯性，這個特性有利於解決切割出的字元具有雜訊或者影像殘缺不全的問題。
本研究將攝影機架於車內後視鏡下方，在實際的市區道路上以正面拍攝前方的車輛作為測試，拍攝環境為白天、晚上、晴天、陰天、雨天。針對前方5~10公尺範圍做偵測。受測對象為一般自用小客車(轎車、休旅車)共257部，以評估系統的效能。

關鍵字

車牌辨識 ； 小波理論 ； 型態學 ； 類神經網路

並列摘要

License plate recognition system has been extensively used in a variety of applications, such as parking lot management of community and buildings, highway toll systems, car management and so on. Most commonly available license plate recognition systems recognize only in static pictures. In this study, a motion-based targets vehicle license plate recognition prototype is proposed. Multiple-frames of pictures are used for recognition. As a result, the time for preprocessing stages is critical and developing time-efficient algorithms become more important. This prototype is designed mainly for specific car-search that should be helpful for many organizations such as police departments and bank.
This paper presents a algorithm for license plate recognition using Discrete Wavelet Transform (DWT) and Neural Network. Classified into four parts: Locating the car-back in an image. Locating the license plate. Segmenting characters of the license plate. Recognizing each character of the license plate.
We accomplish the license-plate localization by high-pass wavelet coefficients. Since the amount of data becomes 1/4 only, we can reduce, a lot of the system required time, the computational complexity, the memory usage, and rid of noise. This methodology provides high efficiency for locating a license plate from an image.
In the recognition, we use the neural network because of the ability that a large number of neurons imitate the neural network of living beings. Besides, neural network has high and getting fault-tolerant, and that is helpful to against the noisy or incomplete cut characters.
In this system, we set video camera under the rear-view mirror and continuously shoot the front motion-based vehicle. The video can be taken under arbitrary weathers and illuminations. In order to test the system efficiency, there are 257 cars been tested.

並列關鍵字

Wavelet transform ； Morphology ； License Plate Recognition ； Neural Network

參考文獻

[18] 梁智凱,“多車牌辨識系統之研究,” 私立淡江大學,電機工程碩士論文,2006.

[11] 陳麗奾, “在未設限環境下車牌的定位與辨識,” 國立台灣師範大學,資訊教育研究所碩士論文,1999.

[15] 王振興,“多標的汽機車車牌辨識系統之研究,” 私立元智大學,資訊管理所碩士論文,2003.

[12] 魏銪志, “動態多標的車牌辨識系統之研究,” 私立元智大學,資訊管理研究所碩士論文,2000.

[8] S. Y. Zhang, M. G. Zhang, and X. Z. Ye, “Car Plate Character Extraction under Complicated Environment,” in IEEE International Conference on Systems, Man, and Cybernetics, vol. 5, Oct. 2004, pp. 4722-4726.