摘要
現今車牌辨識大多是從固定視角甚至是固定光源的情況下進行，雖能達到極高的正確率，但會需要架設定點相機與光源，且只能被動的取得車牌資訊，若是能將車牌辨識系統搬到能夠在空間所有的維度自由移動的無人機上運行，將可以克服上述定點辨識的缺陷。 本篇研究提出了三個深度學習的演算法，分別能夠: 1. 讓無人機辨識出與車輛的相對位置。 2. 偵測畫面中車牌的位置。 3. 辨識車牌號碼。 實際使用時將(1.)的結果用結合自動控制演算法調整無人機，直至能夠楚辨識車牌的位置，然後以(2.)從畫面中擷取出車牌並還原車牌成長方形，最後由(3.)來辨識其車牌號碼。 且訓練資料僅包含少部分真實圖片，因真實圖片取得不易且標註困難，我們建立了合成車輛與車牌訓練圖片的流程，能夠產生極度逼真的結果。大量合成訓練圖片不僅能解決訓練資料標註不易的問題，還能夠避免神經網路過擬合，讓我們的辨識結果更加穩健。 最重要的是，所有的感測器與運算都能夠在飛機上即時完成，因此不受各種無線傳輸的限制，大大增加了系統的靈活度，讓無人機可以在各種不同的環境下進行任務。 最終以Gazebo模擬、室內與室外三種實驗環境證明整個研究的可行性，無人機皆能夠從車輛側面看不到車牌的位置移動到車輛正面，並以90%左右的成功率辨識出車牌號碼。
並列摘要
Nowadays, most of the license plate recognition systems run under stationary viewpoint even stationary light condition. Although they can reach extremely high accuracy, they need to set camera or light source in specific position. Therefore, the images of the license plates can only be obtained passively. If the license plate recognition system can be applied on a UAV, which is able to move freely in 3D space, then we can overcome the issues mentioned above. In this research, we propose three deep learning algorithms to: 1. make the drone estimate the relative position between the car, 2. detect the location of license plate in the image, 3. recognize the license plate number, In the experiment, we use the result of (1.) as position estimation of the car, and apply automatic control algorithm to adjust the position of the UAV, until the UAV can recognize the license plate number clearly. Finally, the license plate will be reverted into rectangle by (2.), and recognize the license plate number by (3.). What’s more, our training dataset only contains small number of real pictures. Because of the difficulty of collecting the training dataset and labeling, we construct a pipeline of training data synthesis, which can generate extremely realistic images of vehicles and license plates. Huge amount of training data can not only solve the problem of data labeling and collecting, avoid neural network to overfitting, but also make our recognition result more robust. The most important is, all the sensors and algorithms are running onboard and in real time to avoid the restriction of wireless connection, which increase the flexibility of the system greatly, and make the UAV able to complete the mission in different situations. The feasibility of this research is proved by the Gazebo simulation, indoor and outdoor experiments. The UAV can move from one side of the vehicle, where it can not capture the license plate, to the front of the vehicle, and the license plate recognition rate of success is about 90%.