應用於停車場之動態車牌定位系統

Abstract

近年來，由於商業的發展與交通的便利，人們對於車輛的需求增多；隨著車輛數目不斷上升，卻引發了一連串的交通問題。既然車輛數目如此繁多，失竊案件也頻有所聞，而停車場為容易發生竊盜地點之一。因此，如何利用電腦化系統取代人力，來做好停車場的管理是當前最重要的課題。 本文提出一套適用於停車場且能夠動態偵測出車牌的方法，包含「移動物偵測模組」、「車牌定位模組」兩大子系統，目的為解決在光線不穩定的環境條件下，不易偵測車牌的情形。系統首先利用移動物偵測模組中的跳躍式背景相減法，將影像序列編號奇數的圖框(Frame)兩兩相減；考慮車牌出現的幾何位置之後，把疑似車牌所在區塊座標，標記在編號偶數張的圖框上；接著，再利用車牌定位模組裡的掃描灰階變異取得次候選區，透過Sobel的垂直邊緣偵測來保留疑似車牌字元的部份區塊。最後，再利用「佈線演算法」和搭配車牌特徵，找出更精確的車牌位置。 為了驗證此方法，測試的資料庫包括了室外各種天氣如晴天、陰天與雨天以及室內的動態車輛影片。地點為台灣師大的地下停車場及大安森林公園停車場。經由實驗的結果，車牌的正確偵測率高達91.07%，平均處理每張圖框的時間為191ms。

In recent years, due to the development of the business and the convenience of the traffic, people require more and more cars. As the increase of cars, it brings a series of traffic problems. Since there are numerous cars, the cases of larceny come up frequently, and the parking lots is one of the places that the thefts easier happen. Hence, how to utilize computer systems to replace manpower to manage parking lots well is a very important topic at present. This paper has proposed a dynamic license plate detection approach that suits parking lots, including two sub-systems, “Motion Detection Module” and “License Plate Localization Module.” The objective is to solve the problem that license plates are not easy to be detected under the uneven lighting conditions. First, the system uses the jump-background-subtraction of the Motion Detection Module to subtract both odd frames in the image sequences. To consider the geometric location that license plates appear, we determine the suspect location coordinates and label on the even frames. Then, to use the scanning-gray-level-variation inside License Plate Localization Module to get the second candidates, next, make use of the vertical Sobel edge detection to reverse the characters area of suspect license plate. At last, to exploit “Line-Arrangement Algorithm” and license plate features to find out the precise license plate location. In order to demonstrate this approach, the test data includes indoor and outdoor dynamic films. (The outdoor weather includes sunny, cloudy, and rainy). And the experimental places are the underground parking lots of NTNU and the parking lots at the Da-An park. Via the result of the experiments, the accurate rate of detecting license plates is high as 91.07%, and the average of processing time is 191 ms per frame.