應用於機器人之三維重建與導航開發

Abstract

本論文探討了利用RGB-D 相機改善同時定位與地圖構建（SLAM）、三維重建 和移動機器人導航的技術。借用此相機運行ORB-SLAM3 可提供自駕車或機器人 有精準的定位性能，我們透過在校園內的實驗，藉由再次拜訪的景物來優化軌跡以 降低誤差。這樣的回環檢測也被用於加強GMapping 的性能，我們使用二維光達 （LiDAR）與單目相機，通過結合視覺詞袋的概念來優化 GMapping 的結果，以實 現循環檢測和姿勢校正。

除了二維情境，我們還使用RGB-D 相機實作SurfelMeshing，以網格（Mesh） 來重建校園中常見的場景。我們不僅改善軌跡精準度，更實現即時的三維重建。雖 然大多數的三維導航任務需要密集地圖，但由多個平面三角形組成的網格可以簡 化機器人導航中的計算。我們產出的網格重建可進一步建立導航網格（NavMesh）， 由於不同的地圖格式，可以在三維空間中導航機器人。導航網格不同於二維網格的 是，它可建立不同高度樓層的可行走區域。導航網格中的多邊形，因大小形狀等更 加彈性的個性使其比普通網格更能精確表示任意環境。

去曾提出的路徑規劃（包括DAO*和DDAO*）為二維網格地圖上的機器人提供了有效的導航策略，即便是存在動態障礙物亦能完成導航任務。我們更近一步改善其迭代過程，以減少每次路徑搜索的計算時間，得到更有效率的導航計算。而導航網格為上述得到的三維重建提供易於理解的地圖方案，我們比較了各種路徑規劃方法，在即使有額外障礙物的三維空間中也能成功完成機器人導航。

本研究的應用包括增強交通、物流和製造等行業中自主車輛和機器人的能力。也因本研究有精確定位與好的三維重建結果，亦有在虛擬現實與增強現實發展的潛力。總體而言，本研究有潛力在未來開發更先進的移動機器人和自主系統的發展。

This thesis investigates visual techniques for enhancing Simultaneous Localization and Mapping (SLAM), three-dimensional (3D) reconstruction, and navigation in mobile robots utilizing an RGB-D camera. The RGB-D module in ORB-SLAM3 delivers precise performance for autonomous vehicles and robotics, with minimal drift due to loop-closing detection, as demonstrated by our experiments on campus. We evaluate and augment prior work, such as GMapping, which accurately localizes in a two-dimensional (2D) grid map but fails to address loop closure. By incorporating the concept of a visual bag of words and combining a 2D LiDAR sensor with a monocular camera, we enable loop detection and pose correction.

In addition to 2D scenarios, we focus on colorful mesh reconstruction from RGB-D cameras and develop a real-time pipeline for 3D reconstruction. By integrating ORBSLAM3 with SurfelMeshing, we determine not only the camera pose but also reconstruct the 3D environment. While most 3D navigation tasks necessitate a dense map, meshes composed of multiple 2D triangles simplify calculations in robotic navigation. Our mesh reconstruction facilitates the creation of navigation meshes (NavMesh), which can be employed to navigate a robot in 3D space due to the distinct map format. Unlike a 2D grid, it permits walkable areas that overlap above and below at different heights. Polygons of varying sizes and shapes in the form of NavMesh can represent arbitrary environments with greater accuracy than regular grids.

Prior work on DAO* and DDAO* provides effective navigation policies for robots on 2D grid maps, even in the presence of dynamic objects. However, the iteration process is inefficient. We have improved the iteration process of DAO* to reduce computation time for each path search for 2D navigation. Besides, NavMesh offers a comprehensible solution to 3D mapping, and we compare various pathfinding methods to demonstrate successful robot navigation in 3D space, even in the presence of additional obstacles.

The potential applications of this research include enhancing the capabilities of autonomous vehicles and robots in industries such as transportation, logistics, and manufacturing. It could also have implications for virtual and augmented reality, where accurate and efficient 3D mapping is essential. Overall, this research has the potential to contribute to the development of more advanced and capable mobile robots and autonomous systems.