高維影像特徵於物件追蹤之研究

Abstract

物件追蹤的技術是電腦視覺領域中一個相當重要的課題，而透過擷取影像中的局部特徵，並從中計算特徵描述，以此特徵描述實現的物件追蹤，可以達到穩健且快速。但由於特徵描述通常都是高維度的，因此需要設計不同的策略進行特徵描述搜索，以加速整體的追蹤效率。 在本研究中，分別介紹能夠偵測影像中尺度不變、旋轉不變的特徵，並生成64個維度特徵描述的SURF演算法，以及搭配均值聚類樹的優先搜索演算法，最後結合此二個演算法，實現一套物件追蹤系統，證實了相較於傳統的線性搜尋法，使用優先搜索演算法能夠使物件追蹤的每秒幀率上升約30%。

Object tracking is a challenging topic in computer vision. Local descriptors generated from distinctive image features can perform a robust and rapid object tracking. However, the feature descriptors are usually vectors in a high-dimensional space, causing the conventional linear search method inefficient. In this research, two algorithms are introduced. The first one is called “SURF”, which is designed to detect the scale-and-rotation-invariant feature in a image, and then generate descriptors for each feature point. The second one is the priority search method combined with the K-means tree. This algorithm can perform fast searching especially for high-dimensional data such as SURF descriptors. In the experiment part of this study, it is showed that the priority search method can raise the object tracking FPS for about 30%, in contrast to conventional linear search.