利用輪廓資訊辨識物體之方法發展

Abstract

物體辨識這個議題近年來在電腦視覺(computer vision)上引起了不少的討論。因為這種資訊可以用在非常多方面像是：機器人學、影像搜索、人機互動等等方面。因為這還是一個未解決的問題，所以如何辨識物體的方法有很多，在本文之中，我們用輪廓資訊來當我們偵測物體的模型。 物體辨識是一個非常不容易的問題，因為我們最後的結果必須是知道物體在整張圖片的位置與大小。但是因為這些資訊不會事先就知道，因此.如何從整張影像中找到物體是一個非常不容易的事情。這也就是說在整個圖片搜尋的過程中，辨識的方法必須從大部分錯誤的圖片之中找到極少正確的圖片。 雖然其他資訊也可能很重要，但本論文的研究著重於只用輪廓的資訊偵測物體。整體而言可以分成兩大部分：第一部分是如何從使用者給予的圖片中擷取出輪廓。這個問題著重在怎麼開始，怎麼形成一條可用的輪廓。我們用了一些找興趣點(interesting point)的方法，像是Fast corner detector與 Difference of Gaussian 等方法。而找出線斷的方法，我們使用的是現在被廣泛使用的Canny edge Detector。有了這兩種方法後，我們可以找出輪廓的基礎模型。而第二部份則是決定於給予的圖片。若是給一張乾淨的影像，我們即可用輪廓來偵測物體。但是也可以給一些正確的圖片，我們也可以運用一些機器學習(Machine Learning)的方法，讓電腦自動找出重要的輪廓線，形成輪廓模型。 最後在與影像比對上，我們用了一種快速比對的方法：Chamfer Matching。在比對每一張圖片的相似度後，我們會有一個物體機率分布圖，我們找到分部最高的點，來當作我們的偵測結果。

Object detection has become an interesting topic in the filed of computer vision. Object information is very important and can be used in robotics, image retrieval and other technical areas. Object detection is a very challenging problem in computer vision, because the position and size of objects in an image must be known by computer itself. When the position and size of objects are not known, the detector should be able to differentiate between the query object and the negative objects. Although many methods and research papers have been presented about how to detect objects, we have chosen a native way to detect objects: contour-based object detection. The purpose of this thesis is to present one kind of object detector using only contour information. The creation of such a detector could be separated into two parts. Firstly, those contours of the object must be formed and found. We use the result of some interesting point detectors as a starting point, and combine them with a well-known edge detector – “Canny edge detector”. Thus, we can identify the object contours in the image. Secondly, our object detector is formed by inputting one image or inputting a series of positive images. We must deal with them in different ways. If a pure picture is inputted with a specify object, we can only matching it without any training methods. But if several pieces of data are given, we try to use some learning methods such as SVM (support vector machine) to learn the model and classify the object in the image. Finally, we use the method of “sliding windows” to detect objects in testing stage. We match the model with the query sub-image. By using “Chamfer matching” we can get a probability map of the object. The highest point is the object center detected by our detector. The benefit of the Chamfer matching is that it can efficiently match edges, which suits our purposes. We compare and show the results in different methods and objects.