Yi dong ji qi ren de zi yun dong jiao fu zhu shi jiao zhuang tai gu ji

Ph.D.State estimation of mobile robots, including localization, mapping and calibration, is a fundamental task for their autonomous navigation. With the huge application potential of image sensors, many impressive vision-based state estimation approaches have been presented over the past 10 years. However, accurately and robustly localizing a mobile robot in unknown environments using vision remains challenging. To overcome the inherent limitations of monocular vision, such as unrecoverable scale and tracking loss, proprioceptive sensors, including wheel odometers and inertial measurement units (IMU), can be introduced to aid visual estimation via sensor fusion.This thesis aims to develop accurate and robust visual state estimation schemes aided by odometry/IMU for ground mobile robots, by investigating to leverage the special motion prior constraints. For the pose parameterization of ground vehicles, conventional SE(3) or SE(2) pose representations are either under-constrained or over-constrained, hence are relatively inaccurate in motion estimation, and not robust enough to abnormal measurements or out-of-constraint motion perturbations. This thesis proposes an approach that is distinct from the conventional parameterization methods. In indoor or small-scale outdoor scenes, the vehicle pose can be parameterized as the SE(2)-constrained SE(3) manifold; the out-of-SE(2) perturbations can be applied on the pose, or be incorporated into the measurement model of the landmarks.Utilizing the new state parametrization, estimation systems are built with different sensor combinations. An SE(2)-constrained state estimation scheme using monocular vision and odometry is proposed for ground vehicles in indoor industrial environments. When using vision and IMU, an SE(2)-constrained visual inertial fusion scheme is presented for motion estimation of ground vehicles in small-scale indoor/outdoor scenarios. An improved approach of the SE(2)-constrained model is also proposed by incorporating the out-of-SE(2) perturbations into the landmark measurement model, and is implemented in a visual-odometric simultaneously localization and mapping (SLAM) system. In all the presented approaches, on-manifold nonlinear graph optimization techniques are employed for accurate state estimation, and the geometric prior constraints are incorporated in the systems via on-manifold operations on the state variables. Experimental tests were conducted on all the systems, to validate their better accuracy and robustness than the state-of-the-art visual or visual inertial estimation systems, without sacrificing the computational efficiency. Meanwhile, although experiments are presented only on ground vehicles in this thesis, the proposed SE(2)-constrained model is generic and potentially applicable to other robotic platforms and scenes, e.g. riverine mapping and navigation.移動機器人的狀態估計，包擴定位、建圖和標定，是實現其自主導航的基本問題。由於視覺傳感器在信息豐富度上的巨大潛力，過去十餘年間出現了很多基於視覺的狀態估計方法，其中不乏令人印象深刻之作。不過，未知環境中移動機器人基於視覺的精確魯棒定位依然是個充滿挑戰的課題。為了彌補單目視覺缺失真實尺度和易於跟蹤丟失的固有缺陷，使用「自運動覺」（Proprioception）傳感器，如輪式里程計和慣性測量單元（IMU），將其與視覺融合，是一種較好的輔助手段。本文旨在通過融合單目視覺和里程計或IMU，為移動機器人構建精確魯棒的狀態估計系統。特別地，本文主要關注地面車輛的狀態估計，探索和利用地面運動中的特殊幾何約束，來提高狀態估計算法的性能。對於地面車輛的位姿參數化，傳統三維或二維空間內參數化失之於欠約束或過約束，因而運動估計的準確度相對較低，且對異常測量和運動約束外擾動的魯棒性欠佳。本文提出一種不同於傳統參數化的新方法。在室內或小規模室外場景中，車輛位姿可以參數化為受SE(2)約束的SE(3)流形。在建模時，SE(2)之外的三個自由度的擾動可以施加於位姿本身，也可以歸併至對路標點的觀測模型。利用上述狀態參數化新方法，本文構建了幾種使用不同傳感器組合的狀態估計系統。針對室內工業環境，構建了一個基於單目視覺和里程計、受SE(2)約束的地面車輛估計方案；在使用視覺和慣性傳感器時，針對室內或小型室外場景，構建了一個受SE(2)約束的地面車輛視覺慣性運動估計方法；同時，提出了一個將SE(2)外擾動歸併至路標點觀測模型的改進模型，并将該模型在一個基於視覺和里程計的同步定位與建圖（SLAM）系統上實現。上述所有系統均使用流形上優化方法來實現精確估計，上述幾何約束則藉助流形上的操作作用於狀態變量上。我們對上述系統進行實驗測試，驗證了其精度和魯棒性相比主流前沿視覺估計或視覺慣性估計系統均有所提升，且不犧牲計算效率。與此同時，儘管本文僅展示了地面車輛上的實驗，但受SE(2)約束的SE(3)流形是一個通用模型，可應用於其他機器平台和場景，如河面的導航和建圖。Zheng, Fan.Thesis Ph.D. Chinese University of Hong Kong 2019.Includes bibliographical references (leaves 118-126).Abstracts also in Chinese.Title from PDF title page (viewed on 29, December, 2020).Zheng, Fan