單攝影機仿人類雙眼視覺在未知環境中之自主定位與運動姿態之估測

The Measurement of Motion and Self-Location with Monocular of Humanoid Binocular Vision in an Unknown Environment

賴璿光(Shiuan-Guang Lai)

指導教授 ： 陳傳生

元智大學/工程學院/機械工程學系/碩士(2012年)

https://doi.org/10.6838/YZU.2012.00060

摘要

機器視覺長久以來一直致力於結合影像處理與自主定位，提供移動載具強健與明確的運動行為，以達到最佳自動化控制的應用。機器能在整體環境中定位自身的位置座標與運動姿態進而做出追蹤目標物、分析目標物的運動姿態，或是建構全環境地圖更是當今機器視覺領域中的首要步驟。

為達到即時定位的目標，在特徵點擷取( Feature Extraction )上，本文選用目前擷取特徵點速度最快的SURF( Speeded-Up Robust Features )特徵擷取演算法，來做為最適合於本文辨識整個Frame特徵點的擷取與匹配，當作定位前的影像處理。接著利用匹配點( Matching Pair )，搭配基礎矩陣( Fundamental Matrix )便可得知自身攝影機相對於整個空間的旋轉角度與運動姿態。

本論文致力於整體的處理速度與精準度，在避免減少特徵點而造成匹配不完全與去除錯誤匹配點上，取得了平衡的實驗結果。本論文的另一個特色在於程式上，可同時匹配四種不同的影像輸入( image-to-image, video-to-webcam, video-to-itself, video-to-other-video )，當想要隨時更換欲研究的影像時，可在程式上做離線處理，不用再每次Compile，非常方便。

關鍵字

SURF ； 機器視覺 ； 影像匹配 ； 影像幾何 ； F-matrix ； 視覺定位

並列摘要

In order to obtain the best application of automatization control, vision machine has been striving to provide mobile vehicle with robust and accurate moving actions by combining image process and auto-localization. The initial step of visual-based machine is to enable the machine to locate its coordinate and motion so as to track object, analyze poses of objects, and construct the map of overall environment.

To achieve targets located in real time, this thesis chooses the fastest method of extraction feature - SURF ( Speeded-Up Robust Features ), which is the most suitable way to recognize feature extraction and mapping of the whole frame, to extract features as an image process of pre-localization. Subsequently, by using matching pair and fundamental matrix, we can know rotation matrix and translation matrix of CCD relative to the space.

Owing to the emphasis on the process speed and accuracy, the thesis acquires an balance result: avoiding decreased features from causing incomplete matching and removing wrong matching points. The other characteristic of this thesis is that whenever we change a researching frame, we can, at the same time, match four different types of image input ( image-to-image, video-to-webcam, video-to-itself, video-to-other-video, )and always process it off-line in the C++ program without compiling, which is really convenient.

並列關鍵字

SURF ； Vision Machine ； Matching ； Image Geometry ； F-Matrix ； Location

參考文獻

1. C. Harris and M. Stephens, " A combined corner and edge detector ", 1988

2. Lowe D. G., " Distinctive Image Features from Scale Invariant Keypoints ", 2004

3. Ke Y. and Sukthankar R., " PCA-SIFT : A More Distinctive Representation for Local Image Descriptors ", 2004

5. Seongjoon Baek and Koeng-Mo Sung, " Fast k-nearest-neighbour search algorithm for nonparametric classification ", 2002

6. F. Jurie and M. Dhome, " Real time robust template matching ", 2002