利用等位函數法結合形狀資訊做影像序列的物體追踪

Abstract

物體追踪在電腦視覺、監測系統、交通監測站等，都有很大的應用，並且提供了高階影像處理的前置工作。本論文提出一個對影像序列的物體輪廓線追踪方法。主要是結合物體形狀資訊來改進Abdol-Reza Mansouri的方法。 在Abdol-Reza的方法中，追踪演算法主要是透過最小化一個由貝氏定理所推導出的能量函數，來變動所要用來追踪的輪廓線，使其往目標物逼進。此方法有三個優點。第一，不需要做物體運動估測。第二，物體允許任意變形，包括拓樸的改變。第三，物體的輪廓梯度不需要很大。而它所面臨的缺點是，當物體的附近有跟其相近的背景時，追踪的結果將會受到很大的影響。我們希望透過結合物體的形狀資訊來改進這個缺點。當被追踪物體有部分跟背景相似時，明確的部分由原來的方法追踪，並將其結果與形狀資訊對齊，接著由形狀資訊將不明確的部分補足。如果物體本身就整個能與背景分開時，形狀資訊的權重將會自動被付予很小的值，以保有物體允許隨意變形的特性。這樣的整合方式，也將使我們的方法更具有彈性。

Detecting and tracking moving objects has a wide variety of applications in computer vision such as computer vision, video surveillance, traffic monitoring, etc. Additionally, it provides input to higher level vision tasks. This thesis presents an approach to tracking a moving object over a sequence of images. In particular, we improve the Abdol-Reza’s model by coupling with shape prior knowledge for shape perseverance in case of ambiguity. In the model of Abdol-Reza, tracking is achieved by evolving the contour from frame to frame by minimizing an energy functional evaluated by Bayesian theory. There are three two favorable features in this model. First, no motion field or parameters needed to be computed. Second, deformable shapes of the object are allowed and the topology of the boundary is not constrained. Third, no assumption is made on the strength of the edge gradient. However, it also suffers from the constraints imposed on a degree of dissimilarity between the object and the background. A background region similar to the object might corrupt the contour evolution. We want to overcome this drawback by coupling with a shape prior in the associated energy functional. When the object is partially involved in a similar background, the original tracking term in the functional will dominate the result in the unambiguous background part, and the prior shape will guide the object movement in the ambiguous part. If the object is entirely distinguishable from the background, the weight of the shape prior is set low and thus allowing free deformation of the object. Compared to other tracking methods embedded with shape priors, the presented approach is more flexible, retaining the advantage of suffering little constraints on the deformable shape of the tracked object in many cases.