室內場景特徵點對應之改善

Abstract

為了兼顧機器人視覺系統建構室內場景所需的正確性和即時性，本篇論文提出一個演算法來改善特徵點對應的問題，並利用極線幾何的限制刪除錯誤的特徵點對應。我們使用Harris演算法擷取一對立體影像中的角點作為特徵點，接著分析在Harris計算過程所得到的影像梯度，將影像梯度方向量化為九個區域，分別以不同的顏色來代表這些區域，最後得到的彩色碼即可做為每個角點具有的特徵。若左右影像特徵點的彩色碼越接近，則表示此兩點對應到真實場景中同一點的機率越大，但相鄰的區域也可能有彩色碼完全相同的點，為了解決模稜兩可的問題，透過距離、角度限制和雙向對應一致性的要求，可以計算出一個評量對應關係的分數當作篩選標準，如此即可找出初步的對應，最後再加上極線限制，一組正確的對應應該要滿足兩個極線限制：也就是特徵點必須落在對應的極線上、所有的極線必須交於極點之上。觀察極線分佈的情形後，分別對極線距離以及極點到極線的距離設合理的門檻值，如此可去除錯誤的對應。同時以剩下的點計算基本矩陣，並重複以上步驟。我們可以找出最正確的特徵點對應。

To achieve the correctness of the vision system, this thesis proposes an algorithm to solve the problem of feature point correspondences which uses the epipolar constraints to delete false matches. We first extract corner points from a pair of stereo images as feature points by the Harris corner detector. The local gradient directions are represented as different colors for each feature point to form a color code. The point correspondences are then obtained by comparing the color codes as well as the spatial relationships between neighboring feature points. Finally, we add the epipolar constraints to find the best point matches. A good feature point must satisfy two epipolar constraints. One is that each feature point should lie on its epipolar line. The other is that epipolar lines should intersect at the epipole. For a pair of feature points, we compute the epipolar distance and the distance between the epipolar line to epipole. Then, we set two thresholds of distance to discard false matches and recompute the fundamental matrix in iteration until all matches satisfy the epipolar constraints.