【導(dǎo)讀】的幾個(gè)重要組成部分之一。然而許多類(lèi)似系統(tǒng)需要復(fù)雜的視頻采集硬件并且需要與紅外閃光。我們也探討了汽車(chē)制造和模式識(shí)別問(wèn)題，其目的在于搜尋用于部分車(chē)牌。號(hào)碼監(jiān)控并與錄像檔案館聯(lián)合一些汽車(chē)視覺(jué)描述系統(tǒng)。設(shè)施寶貴的信息，并提供以各種情境為執(zhí)法對(duì)象的信息。車(chē)牌識(shí)別問(wèn)題被廣泛認(rèn)為是與許多系統(tǒng)急待解決的問(wèn)題之一。雖然向公眾發(fā)布了一些關(guān)于商業(yè)的準(zhǔn)確性細(xì)。節(jié)，但是部署的車(chē)牌識(shí)別系統(tǒng)僅僅在可操作的條件下才能正常工作。因而，他們有兩個(gè)主要。其次，具有一定性質(zhì)的LPR系統(tǒng)可以當(dāng)作是汽車(chē)的指紋車(chē)牌。換句話說(shuō)，確定車(chē)輛的身。份完全基于附帶的車(chē)牌。的汽車(chē)被調(diào)換的情況，在這種情況下，這些系統(tǒng)將無(wú)法發(fā)現(xiàn)這一問(wèn)題。們對(duì)類(lèi)似的車(chē)輛監(jiān)控系統(tǒng)和檢索時(shí)，該車(chē)輛失蹤與當(dāng)時(shí)存檔的錄像資料以及時(shí)間記號(hào)。檢測(cè)的的車(chē)牌圖像的尺寸范圍從。最常見(jiàn)的習(xí)慣做法時(shí)用現(xiàn)有的LPR光學(xué)。板與人物之間距離等措施。圖像分割成60張圖片集測(cè)試集。

　　

【正文】 ed by AdaBoost used raw pixel intensities, however, probably because of their poor discriminating ability with respect to wide illumination differences. Each weak classifier was a Bayes classifier, trained on a single feature by forming class conditional densities (CCD) from the training examples. When making a decision, regions where the license plate CCD is larger than the nonlicense plate CCD are classified as license plate and viceversa, instead of using a simple onedimensional threshold. AdaBoost Training In its original form, AdaBoost is used to boost the classification accuracy of a single classifier, such as a perceptron, by bining a set of classification functions to form a strong classifier. As applied to this project, AdaBoost is used to select a bination of weak classifiers to form a strong classifier. The weak classifiers are called weak because they only need to be correct 51% of the time. At the start of training, each training example (xi, yi) is assigned a weight wi = 1 2m for negatives and wi = 1 2l for positives, where y 2 {0, 1}, m is the number of negatives, and l is the number of positives. The uneven initial distribution of weights leads to the name ”Asymmetric AdaBoost” for this boosting technique. Then, for t = 1, ..., T rounds, each weak classi Pfier hj is trained and its error is putedd as _t = i wi |hj(xi) ? yi|. The hj with lowest error is selected, and the weights are updated according to: if xi is classified correctly, and not modified if classified incorrectly. This essentially forces the weak classifiers to concentrate on ”harder” examples that are most often misclassified. After T rounds, T weak classifiers are selected and the strong classifier makes classifications according to where, _t = ln In addition to the 359 manually cropped positive training examples, we generated additional positive examples by extracting images from 10 random offsets (up to 1/8 of the width and 1/4 of the height of license plates) of each license plate location (for a total of 3,590). We found that this yielded better results than just using the license plate location for a single positive example per handlabeled region Of course, when the detector was in operation, it fired at many regions around a license plate, which we in fact used as an indication of the quality of a generate negative examples, we picked 28 license platesized images from random regions known not to contain license plates in each positive frame, which resulted in 10,052 per set. We then applied a sequence of two bootstrap operations where false positives obtained from testing on the training data were used as additional negative examples for retraining the cascade, and obtained 9,948 additional negative examples. Results Scanning every possible location of every frame would be very slow were it not for two key optimization techniques introduced by Viola and Jones – integral images and cascaded classifiers [26]. The integral image technique allows Detection Rate (%) Figure 3: ROC curve for a 6stage cascaded detector, with 2, 3, 6, 12, 40, and 60 features per stage respectively. Figure 4: Examples of regions incorrectly labeled as license plates. for an efficient implementation and the cascaded classifiers greatly speed up the detection process, as not all classifiers need be evaluated to rule out most nonlicense plate subregions. With these optimizations in place, the system was able to process 10 frames per second at a resolution of 640 480 pixels. Figure 3 shows a receiver operating characteristic (ROC) curve for our cascaded detector. We did not achieve as low a false positive rate per detection rate on our datasets as either Chen and Yuille, or Viola and Jones, but the false positive rate of % for a detection rate of % in set 3 is quite tolerable. In practice, the number false positives per region of each frame is small pared to the number of detections around a license plate in the frame. Therefore, in our final detector we do not consider a region to contain a license plate unless the number of detections in the region is above a threshold. Figure 4 shows a few examples of regions that our detector incorrectly labeled as license plates in our test dataset. Perhaps not surprisingly, a large number of them are text from advertising on city buses, or the UCSD shuttle. Those that contain taillights can easily be pruned by applying a color threshold. 3 License Plate Recognition In this section, we present a process to recognize the characters on detected license plates. We begin by describing a method for tracking license plates over time and then describe our optical character recognition (OCR) algorithm. Tracking More often than not, the false positive detections from our license plate detector were erratic, and if on the car body, their position was not temporally consistent. We use this fact to our advantage by tracking candidate license plate regions over as many frames as possible. Then, only those regions with a smooth trajectory are deemed valid. The tracking of license plates also yields a sequence of samplings of the license plate, which can be used as input to a superresolution preprocessing step before OCR is performed on them. Numerous tracking algorithms exist that could be applied to our problem. Perhaps the most wellknown an popular is the KanadeLucasTomasi (KLT) tracker [23]. The KLT tracker makes use of a Harris corner detector to detect good features to track in a region of interest (our license plate) and measures the similarity of every frame to the first allowing for an affine transformation. Sullivan et al. [25] make use of a still camera for the purposes of tracking vehicles by defining regions of interest (ROI) chosen to span individual lanes. They initiate tracking when a certain edge characteristic is observed in the ROI and make predictions