基于雙目立體視覺在線測量系統(tǒng)的關(guān)鍵技術(shù)研究

1、摘要 I 摘 要 在幾十年的發(fā)展歷史中，計(jì)算機(jī)視覺經(jīng)歷了從 2 維圖像處理到 2.5 維圖像處理，再到 3 維圖像處理這一過程。隨著計(jì)算機(jī)技術(shù)的迅猛發(fā)展，雙目立體視覺作為一種 3維圖像處理手段，已成為計(jì)算機(jī)視覺的重要研究領(lǐng)域之一。其中，雙目立體視覺測量一直是計(jì)算機(jī)視覺領(lǐng)域的重點(diǎn)和熱點(diǎn)研究課題。雙目立體視覺測量是以視差原理為基礎(chǔ)，通過三角測量法來重建物體三維空間幾何信息的一種測量方法。與傳統(tǒng)的測量方法相比較，雙目視覺測量以其獨(dú)特的非接觸

2、、高精度、快響應(yīng)等特點(diǎn)，在三維建模、質(zhì)量監(jiān)控、目標(biāo)識別、非接觸式面形測量等領(lǐng)域占據(jù)著其他測量方法所無法比擬的地位，因而受到廣泛應(yīng)用。 在學(xué)習(xí)了立體視覺技術(shù)的相關(guān)理論的基礎(chǔ)上，對當(dāng)前國內(nèi)外立體視覺的研究工作進(jìn)行了分析總結(jié)，提出了用基于雙目立體視覺的在線測量系統(tǒng)來替代原項(xiàng)目中單目視覺檢測系統(tǒng)，并著重對攝像機(jī)標(biāo)定和立體匹配這兩項(xiàng)關(guān)鍵技術(shù)展開研究。軟件方面，利用 VC++6.0，結(jié)合 OpenCV 庫和 IPL 庫編程實(shí)現(xiàn)了各模塊中的功能。 本

3、文采用的成像模型是帶畸變的針孔成像模型，主要涉及到的坐標(biāo)系變換包括：世界坐標(biāo)系到攝像機(jī)坐標(biāo)系的變換，攝像機(jī)坐標(biāo)系到成像平面坐標(biāo)系的變換，成像平面坐標(biāo)系到計(jì)算機(jī)數(shù)字圖像坐標(biāo)系的變換。 采用改進(jìn)的張正友平面標(biāo)定法對攝像機(jī)進(jìn)行標(biāo)定。一方面，在張氏方法只考慮徑向畸變的基礎(chǔ)上，改進(jìn)算法進(jìn)一步考慮切向畸變對標(biāo)定的影響，并選用分布在圖像中心附近的點(diǎn)來求取標(biāo)定初始值，從而提高算法的標(biāo)定精度和計(jì)算結(jié)果的抗干擾性。另一方面，對 Harris 角點(diǎn)提取方法進(jìn)

4、行改進(jìn)，提高了角點(diǎn)提取的實(shí)時(shí)性。 在進(jìn)行立體匹配時(shí)，本文選用 Canny 算子提取圖像邊緣特征，首先進(jìn)行初步的灰度相關(guān)匹配，然后計(jì)算匹配點(diǎn)的對極幾何和基礎(chǔ)矩陣，通過 8 點(diǎn)算法、RANSAC 算法以及 LMEDS 算法來進(jìn)一步篩選過濾去除錯(cuò)誤或模糊的匹配點(diǎn)， 得到更魯棒的匹配點(diǎn)，然后利用這些匹配點(diǎn)計(jì)算更精確的對極幾何和基礎(chǔ)矩陣，再通過 LMEDS 算法和外極線來提取新的匹配點(diǎn)，組成最為魯棒的匹配點(diǎn)集，用于恢復(fù)場景的三維信息。 通過攝像機(jī)

5、的外參數(shù)矩陣和求得的匹配點(diǎn)，計(jì)算匹配點(diǎn)的三維坐標(biāo)值，從而得出待測物體的尺寸信息。 關(guān)鍵詞： 雙目視覺測量；攝像機(jī)標(biāo)定；特征提?。涣Ⅲw匹配；三維重建。 Abstract III Abstract After decades of development, the computer stereo vision has become one of the important research fields of computer scienc

6、e. As an important branch of computer vision, binocular stereo vision measurement has been the focus and hotspot in the research of the computer vision. Binocular stereo vision measurement is a measuring method which sim

7、ulates human vision principle. In this method, two cameras obtain images of the surrounding scenery at the same time from different angles (or images can be obtained by single camera from different angles at different mo

8、ments). Then object’s 3D geometry information can be restored by the principle of triangulation, on the basis of the principle of parallax. Thanks to its non-contact, high accuracy, fast response characteristics, binocul

9、ar stereo vision measurement has been widely used in 3D modeling, quality monitoring, target identification and non-contact surface shape measurement. On the analysis and summary of current stereo vision at home and abro

10、ad, this article focuses the research on two key technology of binocular vision, namely camera calibration and stereo matching. Using the open source OpenCV library and IPL library developed by Intel and VC++6.0, the pap

11、er implements the contents of function modules. This article uses the pinhole imaging model with the distortion. Imaging transformation mainly includes the world coordinate system to camera coordinate system transformati

12、on, the camera coordinate system to the imaging plane coordinate system transformation and imaging plane coordinate system to the computer digital image coordinate system transformation. In camera calibration, this paper

13、 uses improved Zhang Zhengyou plane calibration method. On the one hand, comparing to Zhang’s method, the improved method takes tangential distortion into account and chooses angular points distributing near the image ce

14、nter to calculate initial values for calibration. These improvements could help to improve the calibration accuracy and anti-interference of calculation results. On the other hand, this paper used the improved Harris cor

15、ner detection method to improve the real-time performance of the algorithm in the process of corner detection. In stereo matching, this article firstly extracts the edge character by Canny edge detector. Secondly, the ed