擷取及辨識在快車道上的摩托車

Abstract

本論文之研究目的在於提出一個辨識系統來擷取及辨識在快車道上的摩托車。輸入的影像是由一部架設在戶外環境的固定式照相機所拍攝。在這些影像上可能會有許多的車輛，而且在影像上的車牌也會有不同大小的情況，這是由於相機和每一車輛的距離不同所致。這系統可分成三個步驟：偵測移動物體，車牌的擷取和車牌的辨識。 在第一個步驟，提出了一個區塊相減方法來偵測移動物體的方法，這些移動物體有可能是在快車道上的摩托車。為了不用影像上全部的點來完成偵測，首先每張影像被棋盤式的切割成M X N個大小一樣的區塊。根據區塊上兩條對角線所求得的變異數和和相似度，每一個區塊可以被歸類成三個種區塊：低對比、靜態、動態區塊。之後一些歸類錯誤的區塊也可以藉由數學上的型態學方法來更正。 在第二個步驟，提出一個基於投影的方法。這類的方法是在投影的統計量上找兩個綁住車牌的高峰。我們提出一個反向的投影方法來擷取車牌。因為車牌的字和旁邊的背景的對比度高，所以可以藉由邊的量值來找出車牌。邊可以歸類成垂直的和水平的邊，藉由這兩種邊可以產生四種投影：垂直邊的垂直投影，水平邊的垂直投影，垂直邊的水平投影和水平邊的水平投影。因為車牌有許多邊，則一條穿過車牌的掃瞄線在投影統計量上，會是一個高點。所以可以藉由這四個統計投影量，來移除產生低點的掃瞄線。最後，一些錯誤的情況可以由這方法所獲得的字高來加以更正。 在第三個步驟，是要將車牌上的字切開加以辨識。因為車牌可能傾斜，藉由車牌底框可以更正。車牌上每個字的寬度都是小點四倍。所以車牌被切成25等分，每個字就可以切出。辨識核心由三個特徵模組組成：contour directional，crossing count和peripheral background area特徵。最後，一些摩托車車牌上的規則可以用來校正一些錯誤。 在實驗部分，測試了180對的影像。區塊相減的方法有98%的準確率，平均移掉影像上88％的點，而且非常的快。反向的方法有94.4％的準確率。對車牌字的辨識率有90.2％。

In this paper, a recognition system is proposed to extract and recognize license plates of motorcycles on highways. Images are captured by a fixed camera under outdoor environments. The images may have more than one vehicle and different sizes of license plates due to the variation of distances from the camera to the vehicles. The system comprised three stages: moving object detection, license plate extraction and license plate recognition. In the first stage, a block-difference method is proposed to detect moving objects, which are possibly the motorcycles on highways. Without using all pixels in images, each image is tessellated into M X N blocks. According to the variance and the similarity of blocks defined on the two diagonal lines, the blocks are categorized as three kinds: low-contrast, stationary and moving blocks. Some misclassified blocks are corrected by the mathematical morphological method. In the second stage, a projection-based method is used to find two peaks in the projection histogram to bound license plates. A screening method is first used to extract license plates. The edge magnitudes are used to locate license plates since the contrast between the colors of backgrounds and characters in the license plates is sharp. Edges are classified into vertical edges and horizontal edges and form four projection histograms: the vertical projection of vertical edges, the horizontal projection of vertical edges, the vertical projection of horizontal edges and the horizontal projection of horizontal edges. Because license plates have many edge pixels, the count of the scanning line across license plates in the projection histograms will be high. So, some scanning lines with low counts can be removed by using information of the four projection histograms. Finally, the license plates which incorrectly segmented into many blocks can be restored by the character height obtained from the projection profile. In the third stage, character images in the license plates are segmented. Because the license plates may be skew, the skew angle of a license plate is determined from the plate's bottom border. The character widths are the same and are four times of the dot "-" width. The plate region width is divided into 25(6*4+1) units and the characters are cut in the computed locations. The recognition kernel comprises three kinds of features: contour directional, crossing count and peripheral background area features. Finally, the regulations of the license plates of motorcycles are used to correct errors. In our experiments, we tested 180 pairs of images. The block-difference method for moving object detection has 98% success rates, is very fast compared with the previous methods and can remove 88 percent of pixels from an image in average. The screening method for license plate extraction has 94.4% success rates. The recognition rate for characters in license plates is 90.2%.