改良式立體視覺演算法應用於機器人視覺系統之研究

A Study on the Modified Stereo Vision Algorithm for Use in the Robotic Vision System

錢鉦津(Cheng-Chin Chien)

指導教授 ： 陳冠宇

中原大學/工學院/機械工程研究所/博士(2013年)

https://doi.org/10.6840/cycu201300714

摘要

以往立體視覺的研究都假設在精確匹配下固定的攝影機焦距，但在晚近幾年的發展中，可變焦距透鏡系統已經變得非常普遍，並已經逐漸運用於機器人視覺系統上。然而，透過使用可變焦攝影機將會產生了嚴重的立體視覺問題：即是倘若透鏡的焦距是不固定或不是確切已知，則以傳統的立體視覺其深度估算誤差將會隨之增加。加之傳統的各種自動對焦演算法在鏡頭焦距、光圈值及目標物深度三者變數之變動下，並未利用立體視覺推估目標物感測深度，而是透過影像清晰度值漫無目的地去搜尋可能的景深區域，然後逐步驅動數位攝影機的調節機制以完成對焦的動作。此二者將造成機器人視覺系統對周遭環境薄弱的感測能力。
因此，本文的研究有三個主要目標：一即是開發適應可變焦距透鏡系統且更精確的改良式雙目立體視覺演算法，並藉以計算得出物體更精準的真實世界座標；二是藉由改良式雙目立體視覺演算法發展更精準穩定的改良式立體視覺開放空間域自動對焦演算法，進而找出在開放空間域裡鏡頭成像最清楚的對焦點位置；三則是利用前二項之研究成果將其運用在立體視覺影像追蹤系統，藉由本文研究之改良式立體視覺演算法與改良式立體視覺開放空間域自動對焦演算法的影像伺服功能，自主移動型機器人可以更精準及更快速地感測周圍環境，避開障礙物，跟隨移動目標物行動、進行環境的動態監測、或是完成指定任務。
經實驗證明，本文的研究的成果具有六項研究貢獻：第一項是改良式雙目立體視覺演算法可以適應可變焦透鏡系統，並大幅改善傳統雙目立體視覺演算法不準確的深度估算狀況。第二項是非均勻離散深度層級誤差修正式已經大幅提升在相似空間的立體視覺精準度。第三項是仿射幾何空間縮放變形誤差修正式更是已經大幅提升仿射空間的立體視覺精準度，仿射幾何空間縮放變形誤差修正式絕對是目前已知準確度最高的雙目立體視覺演算法。第四項是改良式立體視覺開放空間域自動對焦演算法可以適應開放空間域特性，並提升自動對焦系統的精準度及可靠度。第五項則是多重閥值累加方法改良多重閥值顏色辨識法以增進影像追蹤系統在不同環境的追蹤適應力。最後則是，結合上述研究而成的改良立體視覺影像追蹤系統可以穩定而精準的追蹤各種物體。

關鍵字

自對對焦 ； 立體視覺 ； 影像伺服 ； 機器視覺 ； 仿射幾何

並列摘要

Previous studies have stereo vision exact match assumed under a fixed focal length of the camera, but in recent years of development, variable focal length lens system has become very popular, and has been gradually applied to robotic vision system. However, through the use of variable focus cameras will have a serious problem of stereo vision: that is, if the focal length of the lens is not fixed or is not known exactly, places its traditional depth estimation error will increase. Combined with traditional algorithms in various AF lens focal length, aperture value and the target depth changes in the three variables, no estimation using stereo vision sensing target depth, but through the image sharpness measure aimlessly go search for possible depth of field area, and then gradually drive digital video camera focus adjustment mechanism in order to complete the action. This will result in both robotic vision system is weak on the surrounding environment sensing capability.
Therefore, this study has three main objectives: one that is developed to meet variable focus lens system more precise the modified binocular stereo vision algorithms and draw objects in order to calculate more precisely the real world coordinates; Second, by the modified binocular stereo vision algorithm is more accurate and stable development of the modified stereo vision AF algorithm on open space domain, and then find out in open space domain imaging lens focus position most clearly; three is the use of the former two research results, the modified binocular stereo vision algorithm and the modified stereo vision AF algorithm on open space domain, will be used in its stereo vision image tracking system as images servo function, autonomous mobile robots can be more accurate and faster sense the surrounding environment, avoiding obstacles, follow a moving target actions for dynamic monitoring of the environment, or to complete the assigned tasks.
The experiment proved that the results of this study has six research contributions: The first is the modified binocular stereo vision algorithms can be adapted to the variable focal length lens system, and significantly improve the traditional stereo vision algorithms inaccurate depth estimates condition . The second non-uniform spacing of discrete depth level error correction equation has increased dramatically in the similarity space visual accuracy. The third is the affine space warping error correction equation is already significantly improved formal affine space stereo vision precision, the affine space warping error correction equation is definitely the most accurate currently known binocular stereo vision algorithms. The fourth is the modified stereo vision AF algorithm on open space domain can be adapted to open space domain characteristics, and improve the accuracy of autofocus system and reliability. The fifth is a multi-threshold cumulative method improved identification method to enhance the color image tracking system to adapt to different environments track. Finally, the combination of these studies made improvements stereo vision image tracking system can be stable and accurate tracking of various objects.

並列關鍵字

affine geometry ； machine vision ； image servo ； auto-focus ； stereo vision

參考文獻

[6] 許懷根，應用立體視覺於量測微型飛行器的位置和姿態，碩士論文，淡江大學航空太空工程學系，2011。

[7] 錢鉦津，模糊控制之立體視覺影像追蹤系統，碩士論文，中原大學機械工程學系，2005。

[8] 陳冠宇、錢鉦津、張嘉麟、張永鵬，模糊控制之立體視覺影像追蹤系統，先進工程期刊，1期1卷，頁次21-29，2006。

[9] 傅培耕，即時物體追蹤之立體視覺導引自走車，，碩士論文，中原大學機械工程學系，2004。

[10] 陳冠宇、張永鵬、傅培耕，具物體追蹤功能之立體視覺導引自走車，中原學報，33期1卷，頁次29-39，2005。