自動影像分割技術及其於人體偵測與深度估測之應用

Abstract

影像分割技術在醫學影像處理、交通流量監測、人體偵測和多媒體等方面的應用中佔有主要的地位。深度估測也是多媒體與服務機器人應用的一個重要的課題之一。在無法事先取得任何背景資訊的情況下，從各種不同的場景中將人體或其他感興趣的物體抽取出來當成前景，且從單張影像所獲得的資訊估測人體，或者物體的深度將是一件非常具有挑戰性的工作。為了解決這個問題，我們結合了影像中特徵、輪廓和空間分佈等資訊來判別不同的分割區域是否代表著同一個物體，並將判別為同一個物體的不同區域做合併。接著我們使用了兩種不同的深度估測方法對已經從場景中被截取出來的人體作深度的估測。這兩種深度估測方法分別是以消失線及消失點為基礎的估測方式，和以固定相機之查表法的估測方式來重建二維平面影像的三維景深資訊。 在此篇論文中，我們結合了影像分割和人臉偵測技術，希望能在不同場景中將人體抽取出來。首先，我們利用了膚色資訊與橢圓樣板比對找出人臉在影像中的位置，接著我們提出了一套改良式自動種子區域成長演算法來分割影像。在我們提出的分割演算法中，初始種子將會自動產生，且剩餘未分類的像素將會歸到其最接近的區域。在影像初始分割完成後，任兩個相鄰的區域如果具有高相似性將被合併。再根據人體的形態找出人體的範圍，將分割結果屬於人臉與人體的區域作合併，完成前景人物偵測，並確定人體的位置。最後，我們將偵測人體位置的垂直y座標值，並對其做深度估測。假如影像中的消失線或消失點可以被偵測，我們可採用cross-ratio的關係式來估測深度。另一方面，我們藉著建立相機深度查表法，我們可採用查表法來估測其深度。

Image segmentation plays an essential role in applications such as medicine image processing, traffic flow magnitude monitored, human detection, multimedia applications, and many others. Depth estimation is one of the important topics in multimedia and service robot applications. It is a very challenging task to extract human or other objects of interested from scenes without any background information, and then to estimate the human depths from single camera view. To solve this, we adopt the method which combines the feature-based, shape, and space information of an image to recognize different segmented regions. Then we estimate the human depth based on vanishing line and point, or based on camera’s depth look-up table. In the thesis, we combine image segmentation techniques and face detection methods to extract the human from scenes. Firstly, skin regions are detected and an ellipse fitting method is employed to detect the face region and consequently locate the human position. Then we propose an improved automatic seeded region growing algorithm to segment the image. The initial seeds are generated automatically, and the remaining pixels are classified to the nearest region. After the region growing procedure, two neighboring regions with high similarity are merged. The human body is determined by confining semantic human body region in segmented regions, and those belonging to the human face and human body are merged afterward. The human is extracted and the human position is also decided. Lastly, we will detect the human vertical y-coordinate values in the image, and the depths can then be estimated according to the cross-ratio formula or the depth look-up tables of the camera.