請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/51298
標題: | 基於單眼視覺慣性與環境感知之無人載具環境探索與地圖建立 Monocular Visual-Inertial-Based Navigation for Mobile Robots with Perception Aware Exploration and Map Construction |
作者: | Chung-En Liu 劉仲恩 |
指導教授: | 連豊力(Feng-Li Lian) |
關鍵字: | 主動同步定位與地圖建構,單眼視覺同步定位與地圖建構,定位,多解析度分析地圖,感知意識導航,無人載具, Monocular SLAM,active SLAM,localization,multi-resolution map,perception aware navigation,mobile robot, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 本篇論文為無人載具主動同步定位與地圖建構之系統,透過主動找尋最佳視野方向的方式,降低單眼視覺同步定位與地圖建構之定位誤差,並建立多解析度分析地圖。在此系統中,無人載具倚賴單眼視覺慣性達成定位,並將角點偵測影像傳回電腦以求出下一個最佳的視角方向來減少定位誤差。多解析度分析地圖是基於結合定位以及彩色深度攝影機的點雲資訊建立的。本篇論文的貢獻在於,藉由最佳化下一個視角來提升無人載具在無特定結構環境中的定位準確度並建立一個基於周圍物體距離的多解析度分析地圖。 整套系統由六個平行運算的程序組成。VINS-Mono程序藉由彩色影像和慣性測量資訊提供無人載具的定位資訊。RealSense程序負責提供周邊環境的點雲資訊。Octomap程序將定位資訊和不同解析度的點雲資訊結合在一起,以建立多解析度分析地圖。感知意識導航程序藉由求出下一個最佳化的視角,提升無人載具在無特定結構環境中的定位準確度。位置控制器藉由提供基於無人載具和給定終點的距離成比例的速度。偏擺控制器界定可以轉動的範圍並負責提供偏擺的速度。 最後,本論文透過實驗結果來展示此系統的可行性以及性能。 In this thesis, an active monocular SLAM system for a mobile robot is proposed with an aim to minimize pose estimation errors and construct a multi-resolution map of the surrounding environment. In this system, the system uses visual-inertial odometry for localization and further processes the feature detection image to compute the best viewpoint for pose error minimization. The multi-resolution map is constructed by fusing pose estimations with pointcloud from an RGB-D camera. The key contributions of this thesis include improved mobile robot localization in poorly textured environments through viewpoint optimization and detailed environment reconstruction while considering the distance of the surrounding objects. The system consists of six programs running parallel, each acting individually but play an important role in either localization or mapping. The VINS-Mono program localizes the robot by fusing IMU measurements with RGB images. Pointclouds of the surrounding environment, with decreasing resolution for objects further away from the robot, are output from the RealSense program and further combined with VINS-Mono localization for map construction. The Octomap program fuses octomaps of different resolution and probabilities to generate a probabilistic map of the navigation space with respect of object distance and hazardous. The Perception Aware Strategy program equips the system with the ability of minimizing pose errors while navigating in textureless environments. A next-best viewpoint is continuously generated based on the number and position of the detected features. The pose controller program provides the robot with velocities proportionate to the untraveled distances while considering the effects of high velocities on visual-inertial pose estimation. The yaw controller program outputs angular velocities to the robot and bounds the view of the robot with camera parameters to prevent increasing pose estimation errors due to excessive orientation. Analysis on experiments is provided to demonstrate the effectiveness of the proposed system. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/51298 |
DOI: | 10.6342/NTU202002787 |
全文授權: | 有償授權 |
顯示於系所單位: | 電機工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-1008202012513300.pdf 目前未授權公開取用 | 17.65 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。