【作者】 鲁静；

【导师】 杨克成；

【作者基本信息】 华中科技大学 ， 光电信息工程， 2013， 博士

【摘要】 激光水下成像被广泛应用于海洋、江河及湖泊探测等领域，这给基于视觉的水下目标探测提供了可能。利用二维图像信息和三维空间目标之间的映射关系，可以反演目标相对于摄像机的方向和位置，从而得到目标的运动方向、速度和轨迹等信息。圆特征作为目标携带的最常见的特征之一，其定位问题一直是研究的热点。但在水下环境中，光的折射和散射造成了图像畸变和退化，很少有学者针对水下圆目标的定位展开研究。本论文通过连续激光水下成像，建立了水下圆目标的单目视觉成像模型，利用圆目标的单帧图像实现了目标的三维定位。主要的研究内容与成果如下：(1)在室内自然光照环境下通过对直径为500mm圆柱体目标的成像，验证了大气中圆特征的单目视觉成像模型，成功地从单帧图像反演了目标距离，相对误差稳定在2%左右。在大气成像模型的基础上，针对水下摄像机通常被密封在水密容器中拍摄图像的情况，提出了水下圆特征的单目视觉成像模型。该模型对光折射造成的图像畸变进行了补偿，并将目标定位问题分解为求圆锥方程和坐标系变换的问题；(2)基于水下圆特征的单目视觉成像模型，利用空间几何方法推导了圆目标的方位检测算法。该算法每一步计算都基于几何约束，能够提供闭合解，不存在迭代近似。因此只要像面椭圆方程100%准确，就能零误差地反演目标方位，这一点通过数值仿真得到了验证。算法的局限性在于要求圆特征平面和摄像机平面之间的夹角在15范围内。3DsMax仿真实验表明，圆柱体目标从水下0.5~10m的不同高度入射，算法计算的圆特征法矢误差低于1°、位置相对误差低于3%；(3)像面椭圆提取的精度直接决定目标定位的精度。本论文在传统方法的基础上，提出一种曲率与距离共同约束的椭圆自动检测算法。数值仿真结果表明，算法不受迭代初值影响，逼近速度快。仿真图像实验表明，算法对闭合轮廓的识别正确率在80%以上，对重叠轮廓的识别正确率在90%以上，并且对椭圆重叠率不敏感。无论从给定点集、仿真图像还是真实图像的椭圆检测结果来看，算法的正确率、查全率和精度均优于传统方法，并且不受像素点个数的限制，而传统方法在像素点多于50个时精度明显下降；(4)基于前向散射和后向散射的图像退化模型，对水下激光图像的预处理方法进行了研究。同态滤波可以消除乘性噪声，中值滤波对滤除散斑噪声非常有效，将这两种滤波结合起来，可以改善水下图像的质量。主观上从Canny边缘检测的结果来看，噪声对图像边缘的影响降低，圆柱体目标上的椭圆特征得到了有效保留。客观上PSNR和NMSE两项指标也说明了这种预处理方法的降噪能力、对图像细节的保真能力都优于传统的均值滤波和维纳滤波方法；(5)针对一个实际的水域监测问题，设计了水下圆目标定位系统，给出了系统的总体框架设计以及各模块的详细设计。针对系统实现时的摄像机标定、视场盲区、安装板形变等工程问题，给出了相应的解决方案。然后，利用3DsMax提供的虚拟现实技术，对系统进行仿真，在理想条件下说明了系统的可行性、有效性和定位精度。最后，在船池搭建了实验系统，对不同方位的直径320mm，高1.6m的圆柱体目标进行了定位，同时也对图像预处理方法、椭圆检测方法和圆目标定位方法进行了全面检验和分析。更多还原

【Abstract】 Underwater laser imaging is widely used in ocean, river and lake exploration,which makes vision-based object localization possible. Using the mapping relationshipfrom the3D object to the2D image, the orientation and position of the object withrespect to the camera can be calculated. Localization of circular feature, which is one ofthe most common features, is always the research hotspot. However, very few peopleaimed at the localization of underwater circular features (UCFs), since image distortionand degradation occur in underwater environment. Using underwater continuous laserimaging, based on the monocular imaging model for UCFs, this thesis completes objectlocalization by only single image of the UCFs. The main research achievements are asfollows:Firstly, by the images of a cylinder object with a diameter of500mm in indoornatural lighting environment, the monocular imaging model for atmosphere circularfeatures is checked. The distance from the object to the camera is obtained with anaverage relative error of2%. Based on the atmosphere vision model, the monocularimaging model for UCFs is proposed. The model compensates the image distortion bylight refraction, and decomposes the object localization problem into solving theequation of the constructed cone and coordinate system transformation problem.Secondly, based on the monocular imaging model for UCFs, the algorithm ofcomputing the orientation and center coordinates of the UCFs is given. It is based ongeometric constraint, which provides a closed-form solution, and without any iterationapproximation. Therefore, the result of object localization would be absolutely right aslong as the image ellipse equation is absolutely accurate. This is proven by numericalsimulation of the algorithm. In the meanwhile, the algorithm requires that the anglebetween the circular feature plane and the CCD plane is in the range of15. The3DsMax simulation experiment prove that the error for orientation detection is less than1, and that for position detection is lower than3%.Thirdly, the accuracy of localization is directly determined by image ellipsedetection. Based on traditional methods, a novel ellipse detecting algorithm constrained by edge curvature and orthogonal distance is proposed. Numerical results prove that thealgorithm is not affected by iterative initial value and with good approaching rate.Synthetic image experiments prove good performance both with occluded andoverlapping ellipses. No matter from the results of numerical, synthetic image or realimage experiments, all prove that the ellipse detection accuracy, recall ratio, and fittingprecision of the algorithm are superior to traditional methods. Moreover, the fittingresult is not sensitive to the number of edge pixels, when that of traditional methodsdramatically declines with more than50pixels.Fourthly, by the image degradation model of forward and backward scattering, theunderwater laser image preprocessing method is studied. Homomorphic filter caneliminate multiplicative noise, when median filter is very effective to speckle noise.Combining the two kinds of filter, the quality of underwater images can be improved.Subjectively, from the results of Canny edge detection, the method reduces the influenceof the noise to image edge, and effectively maintain the elliptic characteristics of thecylinder object. Objectively, both value of PSNR and NMSE prove that the method issuperior to the traditional linear filter and wiener filter in both noise reduction and detailfidelity abilities.Fifthly, aimed at a water area monitoring issue, a localization system for objects withUCFs is designed. Solutions for some engineering problems, such as camera calibration,blind area of the cameras, and deformation of the mounting plate are given. Botheffectiveness and localization accuracy are checked by virtual reality simulation of thesystem through3DsMax. At last, an experimental system is constructed in a pool with cleanwater to estimate the location of a cylinder object with a diameter of320mm. The accuracyand stability of the image preprocessing method, ellipse detecting method, and objectlocalization method are thoroughly analyzed by moving and relocating the object.更多还原