摘要
在本篇論文中，我們實現了一個跨攝影機車輛辨識系統，此系統主要是利用embedding的概念轉換車輛特徵進而加以辨識，這種embedding出來的特徵向量特性是在跨攝影機辨識時，並不屬於直接匹配而是間接特徵匹配的方式。這種間接特徵匹配的優點在於跨攝影機的車輛辨識時，我們可以有效的克服不同攝影機下，因照射視角的不同所帶來車輛外觀上的差異，以及不同的照射區域所造成顏色亮度的不同。最後我們在實驗上會針對不同狀態進行測試以證明此系統的穩定度。 首先我們先利用背景相減法將影片中所出現的車輛視為前景萃取出車輛前景區域，接著對此區域內的車輛前景作條件式的型態學運算將前景中破碎區域填補起來，接下來針對此車輛位置區域做運算時，我們必須把擷取出來的車輛區域做調整，這樣才能確保車輛行駛到畫面不同區域時，擷取出來的車輛圖片都能夠歸一到同樣的中心。獲得車輛前景圖片之後，edge和color的特徵都會被萃取出來，進而利用各個特徵之間的距離比對產生一組distance feature matrix。Distance feature matrix主要的用途是用於訓練出各個攝影機下的example vehicle，得到example vehicle之後，會再使用一次特徵距離運算，但這次只針對example vehicle做特徵距離運算，最後我們會得到另一種矩陣 – embedding feature matrix。Embedding feature matrix的用途是訓練辨識時所需要的SVM分類器。最後，車輛辨識系統的辨識結果即為SVM分類的輸出結果。在實驗上我們拍攝現實生活中的道路場景去衡量此車輛辨識系統的準確性，經過實驗的結果，可以確定我們的系統可達到相當高的正確率。
並列摘要
In this paper, we implemented a vehicle verification system, which mainly use the concept of embedding to convert the vehicle's feature and then identify this vehicle. In addition to characteristics, the method for matching vehicle objects across cameras in this thesis is indirectly. The advantage of indirectly matching vehicle objects is to overcome the variations of size, aspect, and illumination obtained from cameras. Finally, we will experimentally test for different states to prove the stability of this system. First, we use background subtraction to extract the vehicle (which appear in the video) as the foreground region, and then we used the conditional morphological operations to fill up the broken region which is in the foreground. We do some operations for the vehicle location of this region, however, we must adjust the vehicle region which is be retrieved by us. This way can ensure that when the car drove to different region, the images are retrieved out of the vehicle can be normalized to the same center. After we obtained the vehicles foreground image, we can get the edge and color features, and then we contrast the distance between this features, and we can produce a set of distance feature matrix. Distance feature matrix is primarily intended to train example vehicles. After when we get example vehicle, we used these example vehicle as our basis for robust distance measure operation, and finally we get embedding feature matrix. The embedding feature matrix is used to train the SVM classifier which is used to identify. In the experiment, we take a video in the road scenarios in real life to measure the accuracy of vehicle identification system. According to the experimental results, we can determine that our systems can achieve very high accuracy.