【作者】 刘天阳；

【导师】 王秀坤； 郭禾；

【作者基本信息】 大连理工大学 ， 计算机应用技术， 2010， 博士

【摘要】 工程物的图像监测,是指通过图像技术对工程作业对象的姿态、位置等参数,以及其它外部特征属性进行监测或者估算。传统的工程物监测方法主要是使用全站仪、GPS以及陀螺仪等专用仪器进行测量,费时费力,而且容易受到人员、场地等外界因素的影响。特别是在海洋环境中,由于现场环境的特殊性,导致很多大型工程物的监测问题长期以来始终难以解决。使用图像方法对工程物进行监测,不需要接触到被测对象,效率较高,而且对环境的要求很低。因此对图像监测方法进行研究在工程上具有重要意义。直观上,工程物姿态参数的测算可以通过图象重构的方法来实现。本文首先介绍了两种工程物图像重构方法。第一种方法是激光扫描重构法,通过激光扫描仪和照相机进行工程物的三维重构；另一种方法是多视图重构法,通过两台摄像机在不同角度拍摄的目标图像来对工程物进行三维重构。实验表明,这两种方法能对工程物进行三维重构,但其结果易受到拍摄距离、拍摄角度这样的外部条件影响。为解决工程环境中的图像监测问题,本文重点研究了如下内容：(1)对工程物的图监测方法进行研究,提出一种工程物姿态的单视图测算方法。重点研究了这种测算方法中的误差来源,并给出一种能够提高测算精度的数据归一化方法。模拟实验表明,这种方法具有较好的鲁棒性；图像实验的结果表明,其对于刚体姿态参数中的角度测算误差小于1.4°,位移测算误差约为16m(相对误差小于1%),基本满足工程的需要。在实际应用中,通过这种方法对深海导管架的入水参数进行了有效估算。(2)对平面图像的度量矫正方法进行研究,结合海洋环境中的图像监测实际,提出一种通用的海面图像矫正方法。这种方法利用一个漂浮在海面的充气圆圈对海面图像进行度量矫正,重点研究了海浪波动对矫正结果的影响,并提出一种能够克服海浪波动影响的充气圈姿态筛选方法。通过模拟实验对这种度量矫正方法的鲁棒性进行了验证；图像实验表明,矫正后的海面图像中角度误差小于2°。(3)对工程物的图像识别方法进行研究,提出了一种基于链码技术的图像分类识别方法。这种方法把图像中的物体边缘转化为链码表示,使用分词方法提取出对应的特征编码,再通过混合高斯模型来区分不同的识别对象。图像实验的结果表明,这种方法适合于识别纹理基元形状近似为圆形的纹理图像。更多还原

【Abstract】 In this paper, engineering-object observation is to estimate the attidude, the position and other features of the object in engineering. Traditional measurement methods, such as theodolite, GPS or gyro, may consume much time in engineering, and may be affected by many unexpected factors. Especially, many measurement problems of engineering-object are unresolved in ocean due to its special environment. In these cases, to measure the attitude of engineering-object with image technolgy has the advantages of non-contact, high efficiency and high degree of automation. Therefore, to make research on image measurement is very important in engineering.Generally, the attitude of engineering-object can be computed directly with image reconstruction. The first part in this paper has introduced two object reconstructions:laser scanner reconstruction and multi-view reconstruction. The first one is to reconstruct engineering-object with two laser scanners and a digital camera. The second method is multi-view reconstruction. An image homography estimation which combining both line and point correspondences is introduced in the reconstruction, and the method is used to make image rectification. Experimental results have shown that the method can make an effective reconstruction to a rigid object in a long distance.Based on the improvement of the two image reconstruction method, this paper mainly studies the following problem:(1) Make research on vision based measurement and has proposed a single view estimation method for rigid object attitude measurement. The measurement error is mainly analyzed, and a normalization method is proposed to improve the measurement accuracy. Simulation experiments and image experiments have shown that the method is robust, and the measurement errors of rigid object rotation angle is less than 1.4°and the displacement measurement error is approximately 16m (relative measurement error is less than 1%). This single-view based measurement method has been applied to estimate the launching parameters of an offshore Jacket.(2) Make research on the image metric rectification and has proposed a general metric rectification for ocean surface image. The method utailze a circle float on the ocean surface to compute the metric rectification matrix of ocean surface image. The variation of the normal vanishing point position of the circle support plane is analyized with simulation, and a method to select the proper attitude of the circle is proposed to decrease the compute errors. Experimental data have shown that the rectification is robust, and the angle rectified error is less than 2°.(3) Make research on the image recognition and has proposed a chain code based image recognition method. Firstly, the image of engineering-object is consider as a kind of texture image and the edges of texture unit is extracted and transformed into chain code; secondly, the feature codes of each kind of image texture are obtained with string matching method. At last, Gaussian mixture model is introduced to distinguish the texture with the feature codes and its repeat frequency rate in each image. Experimental results have shown that the method is suit to recognize the circular texture image.更多还原