緊耦合視覺激光傳感器以提升定位與建圖的效能

Abstract

SLAM(Simultaneous Localization And Mapping, 同時定位與建圖)技術是近代機器人系統的 核心技術之一。近年來，基於激光雷達的SLAM系統已經在多種應用場景展現出令人信服的 效果。另一方面，由於相機的低成本，輕量與節能特性，基於相機圖像的視覺SLAM方案在 攜帶式裝置與小型飛行器上仍是主流。然而，兩種方案皆有其侷限性:激光方案容易受到退 化情境的影響，例如在並缺乏形狀特徵的道路上的定位就表現不佳。相機方案則易受到光照 條件與環境中動態物體的影響，使其在臃堵場景下穩定性較差。 本篇工作致力於解決視覺與激光SLAM中的主要挑戰，並藉由緊耦合視覺與激光傳感器量 測的方式來進一步提升SLAM的準確性，效率及穩定性。在激光方面，本篇工作提出了強健 的激光回環檢測方法，它能比常用的視覺方案檢測到更多回環，並藉由分析SLAM圖模型的 方式來保證檢測正確率。為了同時維持SLAM的效率與精確度，本篇工作提出了GLF(General Lidar Factor)來將多個激光傳感器量測進行壓縮，這可以精簡地保存單次激光定位的資訊。 在視覺方面，本篇工作致力解決極端場景下的圖像匹配問題。視覺SLAM的精度依賴於找出 兩張圖像間的正確匹配特徵。雖然現存方法在輸入圖像為高頻且連續序列時基本能保證正確 性，但在有強烈光照，視角及場景組成變化下的圖像匹配仍是一開放性的難題。這也是 SLAM在真實世界應用中能提升效能的關鍵。 在本篇工作的結尾，我們提出了一新穎的方式來結合視覺與激光SLAM系統。現存工作大 多試圖用激光的距離量測來加強視覺SLAM，或者單純將視覺SLAM的運動估計作爲激光 SLAM的輸入使用。本篇工作提出緊耦合的視覺激光SLAM(Tightly-Coupled Visual Lidar SLAM, TVL-SLAM)。在此系統中，視覺與激光SLAM的前端特徵提取和狀態估計是獨立運 作的，但是視覺與激光量測會在後端進行協同優化來估計整個系統的狀態。此系統是基於開 源的ORB-SLAM2視覺系統，及自行研發的激光SLAM系統進行融合。雖然作為基礎的單傳感 器SLAM本身都只有平均水準的表現，但融合後的系統明顯優於現存最先進的SLAM方法， 在自動駕駛的KITTI訓練數據上達到了0.52%的相對誤差，並在線上競賽的測試集達到了 0.56%的相對誤差。

SLAM(Simultaneous Localization And Mapping) is an essential element for modern robotic systems. In recent years, Lidar-based SLAM solutions are already proved to be promising in many practical applications. On the other hand, because of its low cost, light weight and power consumption, camera-based SLAM solution is highly preferred for portable devices and MAV(Micro Aerial Vehicle) platforms. Nevertheless, both solutions have their draw- backs. Lidar-based solution suffers from the shape-degenerate scenarios, e.g. it is easy to fail on long and straight roads with very few geometry features. Camera-based solution is weak to illumination changes and moving objects, and exhibits poor robustness in crowded scenarios. This thesis aims to tackle the major challenges of the visual and lidar SLAM, then further im- prove the accuracy, efficiency and robustness of SLAM system by tightly coupling the visual and lidar measurements. For the lidar side, we proposed a robust lidar loop detection approach, which can detect much more loops than the vision-only methods, and guaranteed the correctness by ana- lyzing the SLAM graph model. To maintain both the efficiency and accurancy for large-scale SLAM tasks, a general lidar factor(GLF) is proposed to compress a bunch of lidar residuals into one resid- ual, while it preserves the information of the original lidar registration. For the vision side, this work aims to solve the image matching problem under extremely high outlier measurements. The visual SLAM accuracy largely depends on finding correct correspondences between two images. Though most state-of-art approaches can guarantee robust matching for consecutive and high-frequency im- ages, the accurate feature matching for images with large viewpoint, illumination and composition differences is still a open question. Which is significant for improving the robustness and accuracy of visual SLAM in the real world. In the end, we investigate a novel way to integrate visual SLAM and lidar SLAM. Instead of enhancing visual odometry via lidar depths or visual odometry as the motion initial guess of lidar odometry, we propose tightly-coupled visual-lidar SLAM (TVL-SLAM), in which the visual and lidar front-end are run independently and which incorporates all of the visual and lidar measurements in the backend optimizations. The visual-lidar SLAM system implemented in this work is based on the open-source ORB-SLAM2 and a lidar SLAM method with average performance, whereas the resulting visual-lidar SLAM clearly outperforms existing visual/lidar SLAM approaches, achieving 0.52% error on KITTI training sequences and 0.56% error on testing sequences.