在移動物體追蹤系統中以最短距離為基礎的初始框架決策技術

Abstract

移動物體追蹤在電腦視覺領域是一項熱門的題目。在不同領域上有許多不同的應用被研發出來，例如監控追蹤系統、擴增實境等等。在移動物體追蹤系統中，利用平均移動追蹤演算法（mean shift tracking）搭配圖像識別（pattern recognition）技術常用作搜尋一個影像畫面中的局部區域，但是平均移動追中演算法在初始畫面中無法得知物體的初始位置，因此無法準確地決定其初始搜尋框架。在這篇論文中，我們提出了決定初始搜尋框架的方法，解除了平均移動演算法的限制。我們亦利用最短距離方法改善特徵匹配，使初始框架的位置更加準確。最後，我們透過不同的實驗結果來展示所提出方法之效能。

Mobile object tracking is an interesting topic in computer vision. Many applications have been developed in many domains, i.e. surveillance tracking systems, augmented reality, etc. In this work, we adopt mean shift tracking scheme with the pattern recognition to carry out localized search on an image frame. However, the mean shift does not know the initial location of the object which is so called search window. We aim to overcome this limitation by giving mean shift the determination of the search window’s location. Besides, to make the location of the search window accurate enough, we also improve the feature matching by adding a shortest distance method. The efficiency of our proposed approach is demonstrated through various experimental results.