On Constructing the Registration Graph of a 3-D Scene Using RGB-D Image Streams

[[abstract]]在使用裝設 RGB-D 攝影機的機器人探索未知環境，最關鍵的問題是如何整合包含在取得的一系列影像中的資訊。由於攝影機的視角能拍攝到的區域有限，所以需進行局部區域的貼合 (Registration)。若將攝影機每一個拍攝位置取得的影像表為一個節點，而每兩張影像的貼合結果視為連結兩個節點的邊，則完成的貼合結果即可表為一個貼合圖 (Registration Graph)。建構 3-D 場景的貼合圖可大幅化簡 3-D 場景的建構問題，其主要的關鍵是決定場景中的任兩節點是否有拍攝到重疊的場景，若有則需執行兩兩貼合的運算並連結兩個節點。在本研究中，我們是使用兩張影像間各自的 SIFT 特徵點的配對數粗略篩選可能有重疊的影像對，但是傳統的比對方法複雜度過高，所以我們提出一個基於視覺詞 (Visual Words，簡稱 VW) 的比對方法來加速運算。首先，我們先對兩張影像間各自的所有特徵點進行向量量化 (Vector Quantization)，以 VW 表示每一個特徵點。若兩個來自不同影像的特徵點都被量化成不同 VW，即表示這兩個特徵不可能配對。所以在配對運算時，只需對被量化成同一個 VW 的特徵點進行比對，計算量可大幅度降低。最後，因為這些特徵配對點裡會有包含錯誤的特徵配對點，所以需要使用 RANSAC 方法濾除錯誤配對點，找到之間較佳的三維資訊轉換關係來進行貼合。實驗結果顯示使用 VW 來對特徵進行特徵比對的方式可以提高運算的效能，並且不會損失太多特徵點配對精確度。[[abstract]]The key problem of using a mobile robot equipped with an RGB-D camera to explore an unknown environment is how to fuse the information contained in the acquired images. Due to the limited field of view of the camera, it is inevitable to register the acquired images. If we represent each image as a node and each pairwise registration result as an edge linking two registered images, then the completed registration results can be expressed as a registration graph. Constructing a registration graph from a series of input images can greatly simplify the 3-D scene reconstruction problem. Notably, the critical issue of registration graph construction is to determine whether a pair of given images are overlapped. If two images are determined to be overlapped, then the second problem is to determine their registration parameters and to add an edge to link those two images. In this work, we use the number of SIFT feature correspondences to select possibly overlapped images. However, the computational complexity of the traditional SIFT feature matching method is too high. Hence, we propose a fast SIFT feature matching algorithm based on the visual word (VW) technique. We first quantize the SIFT features via the vector quantization method with a specified codebook. If two SIFT features are quantized to different VWs, then those two SIFT features are deemed as not matched. Therefore, when matching SIFT features, we only have to consider those features having the same VW and, thus, the computation cost can be greatly reduced. The matched SIFT features computed with the VW approach are further verified with the RANSAC algorithm to remove incorrect matching results and to estimate the registration parameters. Experimental results show that the proposed method can improve the computation speed for 38 times without sacrificing two much matching accuracy.[[note]]碩