3D 虛擬場景已被廣泛應用於許多場合,如虛擬實境與地理資訊系統等應用。本研究致力於發展一套低成本、能自動重建大尺度場景之自主式移動機器人。為達到此目的機器人必須搭載自行開發之環境資訊收集裝置 (environment information collector, EIC) 來收集環境三維空間資訊與影像資訊。並且採用以擴展式卡曼濾波器為基礎的同步定位與地圖建構技術 (simultaneous localization and mapping, SLAM) 來估測機器人自身位置並連結各場景間之關係。將所收集到的深度與影像資訊分別存放於深度影像與環場影像中,在透過兩張影像間之輪廓對應降低傳統雷射與影像間對應的繁雜手續,加快色彩點雲場景的重建。多場景間的接合是利用疊代最接近點演算法 (iterative closest points, ICP) 來完成,但為了避免地形起伏所造成的影響,故考慮機器人六個方向自由度的姿態來進行場景接合。在完成場景接合後仍需透過三角網格處理建立模型表面,並運用材質貼圖來產生高品質的場景模型。本系統提出改良後的 ICP 演算法來提升場景接合之成功率與降低運算時間,平均單步執行時間為7.9秒。場景與場景之平均距離為10.6 公尺,且接合誤差小於8%。若在接合時偵測到封閉迴路,將利用 ELCH (explicit loop closing heuristic) 演算法來加以修正。
3D virtual scenes have been extensively applied in many fields such as virtual reality and geographic information system. The aim of our research is to develop a low-cost autonomous mobile robot which reconstructs large-scale scenes automatically. To achieve this goal, our robot is equipped with an environment information collector which acquires 3D space and image information of the environment. The EKF-based SLAM was applied to estimate the robot position and link different scenes. The range and color information was saved in panoramic images and range images respectively. The contour projection of these two images had replaced the complex matching procedures between laser points and image and speeded up color point cloud scene reconstruction. Multiple scene registration was done by applying the ICP algorithm. However, in order to avoid the effect of bumpy ground, the robot pose considered six degree of freedom (DoF). After the scene registration, the surface model was constructed by triangular mesh processing. Utilizing the texture mapping has produced high quality scene models. Our system proposed a modified ICP algorithm to improve the success rate and running time of scenes registration. The average running time was 7.9 seconds per step and the average distance between each scene was about 10.6 meters. The error in registration was less than 8%. When the loop closure was detected in a scene registration, the ELCH algorithm was applied to amend it.