| 摘要: | 本文研究以往純光達搭配自適應蒙特卡羅定位法(AMCL)的機器人定位方法，並提出了基於機器視覺與光達感知融合的創新機器人定位系統。本創新定位系統將自主移動機器人(AMR)的定位流程拆解為全局定位、位姿追蹤、及綁架問題三大部分，在分析視覺與光達於各個部分的優劣後，依結果選擇各部分使用的感測器，以提升系統整體的可靠度與效能。藉由本定位系統，首先，將基於一對二維標籤(AprilTag)建立的空間座標系做為AMR全局定位的參考基準，解決光達訊號特徵稀缺，且無法快速、直接地辨識AMR位置的缺點，進而完成初始位姿定位。接著，基於標籤計算出的精確初始位置，利用光達訊號高精度、即時性的特性進行位姿追蹤，有效維持AMR的準確定位。最後，在位姿追蹤的過程裡，持續監測AMCL粒子的離散程度，以確保系統估計的位姿準確度。當總粒子變異數超出特定域值時，系統將會觸發全局定位功能進行重定位，解決綁架問題。
透過本文實驗數據結果，得證以下三項貢獻：(1) 相較於純光達定位系統，本定位系統能以較少的演算迭代與計算量完成AMR定位，並且能夠有效且自動地解決純光達定位系統，於幾何特徵重複性高的環境中難以定位的問題；(2) 利用本研究一對標籤的全局定位算法，可以有效解決單標籤定位時，座標轉換誤差的問題；(3)結合基於標籤的路徑點自動生成方法與上述定位系統，能夠簡化及自動化AMR系統前置作業流程，使本系統更具實用性。
;This thesis studied the previous robot positioning method only using LiDAR and Adaptive Monte Carlo Localization (AMCL) and then proposes an innovative robot positioning system based on sensor fusion of computer vision (CV) and LiDAR. The positioning process of the automatic mobile robot (AMR) of this innovative system is divided into three parts: Global Localization, Pose Tracking, and Kidnapping Problem. After analyzing the advantages and disadvantages of CV and LiDAR in each part, it is studied to select appropriate sensors to ensure the overall reliability and high performance of this system. Through the positioning system, firstly, the spatial coordinate system constructed with two-dimensional tags (AprilTag) is applied to be the reference basis for the Global Localization. With this approach, both issues of the feature scarcity from the LiDAR signals and the inability to quickly identify the AMR position can be solved, and then the initial position can be determined. Secondly, based on the initial position, the accuracy of AMR position can be maintained by LiDAR because of its accuracy and real-time data acquisition. Finally, in Pose Tracking, the dispersion degree of AMCL particles, which is used to evaluate the certainty of localization, is continuously monitored. Once the variance of particles exceeds a certain threshold, the re-positioning function which solves the Kidnapping Problem will be triggered.
The results of these experiments prove three contributions. (1) Compared to the pure LiDAR positioning system, this positioning system can complete the AMR positioning with fewer iterations and computer consumptions, and can also effectively and automatically solve the environmental issue of high repeatability of geometric features. (2) The coordinate transformation error in Single Tag Global Localization method can be effectively eliminated by our Two Tags Global Localization method. (3) By combining the automatic waypoint generation method through tags and the above positioning system, the pre-operation process of the AMR system could be simplified and automated, which makes the system more practical. |
