本研究以多顆CCD攝影機實現遠距且廣角之即時三維地圖模型建立系統。相較於使用Kinect (MicrosoftR Co.) 攝影機,此裝置可在白天進行戶外拍攝,且可視距離較長、視角較廣,不僅可擷取到較豐富的特徵,也能減少誤差的累積,並更容易偵測Loop closure。除了使用加速強健特徵 (Speed Up Robust Features, SURF) 找出連續兩張影像之二維對應特徵點,本研究進一步以交叉比對方式做初步過濾,接著藉由隨機樣本一致性 (RANdom SAmple Consensus, RANSAC) 演算法先去除掉二維的錯誤匹配,再結合視差影像的深度資訊,重新執行三維的RANSAC過濾,消除估測距離錯誤的特徵,最後使用所有的inliers計算出最適當的投影矩陣,並獲得新進影像與地圖之間重疊區域,只將不重疊的部份加入地圖,大幅減少需要儲存的資料量,重複上述動作即可建構出目標空間的三維模型,若建完模型後有發生Loop closure問題,則以Graph-SLAM演算法加以修正。
In this study, a long-distance and wide-angle stereo system for real-time 3D model reconstruction was built. Compared with the Kinect (MicrosoftR Co.) camera, this device can work fine in the daytime outdoor, and the visual distance is longer, the viewing angle is wider. It can not only capture rich features, but also reduce the accumulation of errors and make the Loop closure much easier to be detected. In addition to using the SURF features to find two consecutive images of the two-dimensional corresponding feature points, this study further use cross check to do a preliminary filter in order to pick out the most appropriate corresponding map image. Then combine the feature points with the depth of information, using the RANSAC algorithm to remove the error of the corresponding point and calculate the most appropriate projection matrix to project the new image to map coordinates. After the model was reconstructed, and there is the loop closure detected, the algorithm of Graph-SLAM would be amended.