| 摘要: | 由於救援機器人所處在的工作環境通常為未知的複雜地形環境，為了讓救援機器人能在此環境下 進行準確的自主式路徑規劃及移動控制，精確的導航定位及地圖建構功能將成為重要的基礎功能之 一。本計畫的主要目的即在研發一個可應用在複雜環境下之立體視覺測程系統，以提供自主式救援機 器人能夠在救災任務中進行準確的視覺導航及定位功能之用。為了達到此目的，除了必要的視覺感測 技術外，影像處理中的特徵點偵測與匹配、特徵點追蹤、三維重建、相機移動估測以及系統運算效能 等問題也是必須要考量的。在視覺感測技術方面，為了能提供穩定的三維空間座標重建效果，同時又 兼顧成本的考量下，立體視覺系統是個相當合適的選擇。在特徵點偵測與匹配方面，可採用能夠同時 進行特徵點粹取及特徵描述動作的演算法，不但能提高特徵點匹配的處理速度，同時也能加強特徵匹 配點的準確度及強健性。在特徵點追蹤方面，除了可透過移動平台的移動估測結果進行預測外，也可 透過以影像為基礎之特徵點追蹤法進行特徵點追蹤的任務，藉此提高追蹤效能。在三維重建方面，為 了提高重建結果以及移動估測結果的準確性，可透過立體視覺模組的誤差模型，幫助後端估測處理的 資訊融合動作。在相機移動估測方面，由於離群點的移除對於移動估測結果的準確度有極大的影響， 因此在進行移動估測前，必須搭配強健型離群點移除的估測演算法，分離出群內點以及離群點，藉此 提高移動估測結果的準確度。在提高系統運算效能部分，由於立體視覺測程系統的處理過程中將需要 使用大量的運算，若使用傳統的CPU 運算架構將會影響整體系統的即時性。因此，為了維持系統的 運算即時性並提高輸出的更新率，透過CPU 與GPU 整合運算的CPU-GPU 異質運算技術將會是重要 的關鍵技術。藉由以上技術的結合，本計畫將預期發展出一套以CPU-GPU 異質運算技術為基礎之立 體視覺測程系統，用來作為自主式救援機器人之里程計裝置。我們預計以三年時間完成此計畫，第一 年之重點在影像前處理功能以及立體視覺誤差模型的建立，第二年將集中在特徵匹配點粹取與追蹤功 能的實現。第三年則完成三維重建以及相機移動估測之功能，並將系統整合於自主式救援機器人中， 同時進行系統測試與驗證。
Since a rescue robot usually works in an unknown and complex environment, the function of localization and map-building becomes one of the important fundamental functions in order for the robot to perform path-planning and motion control accurately. The purpose of this project is to develop a stereo odometry system for autonomous rescue robots to handle robot navigation and localization problems during rescue tasks in a complex environment. Since the visual odometry tasks require large computations, it is important to design the vision system with high-performance computing power to achieve real-time processing. In addition to the design of vision system, this project also involves the development of feature detection and matching, feature tracking, 3D reconstruction, and camera motion estimation. For the design of vision system, stereo vision will be a good choice of image capture device under the consideration of both cost and 3D reconstruction capability. As for the development of feature detection and matching algorithm, the simultaneous feature extraction and feature description algorithm is preferred in order to improve the performance and robustness of feature matching process. In the design of feature tracking algorithm, in addition to the feature predication method based on the motion information of robot platform, the image-based feature tracking method is also useful for improving the tracking performance. As to the performance improvment of 3D reconstruction and motion estimation, it is suggested to apply the error model of stereo vision module to multi-sensor fusion processing. In the meantime, since the outliers in matching points have a great influence on the motion estimation result, it is important to add an outlier detector into the design of camera motion estimation process to improve the accuracy of motion estimation. Moreover, to achieve real-time computing performance, the CPU-GPU heterogeneous computing technology is helpful to enhance the computation efficiency of overall visual odometry algorithm. A stereo odometry system with real-time computing capability can be constructed by the combination of the techniques discussed above. Such system can be used as an odometer device for an autonomous rescue robot working in a complex environment. We plan to complete this study in three years. The first year will focus on the image preprocessing and stereo vision error modeling. In the second, the emphasis will be on the image-based feature detection, feature matching and feature tracking. The last year will focus on the implementation of 3D reconstruction and camera motion estimation algorithms. The developed stereo odometry system will be integrated into the rescue robot system to evaluate the performance in practical experiments |
