基于球面几何不变量的鱼眼摄像机定位方法研究

With the rapid development of sensor and network technology and the popularization of digital image technology, vision-based pose estimation technique has been applied in robot navigation, virtual reality and visual surveillance fields and soon. Because field of view of the pinhole model camera is narrow, the key feature in the continuous frame is easy to miss. And the accuracy of camera pose is relatively low. Thus, omni-directional camera with 180 degree field of view attracts many researchers attention. Fisheye camera pose estimation based on spherical geometry invariants and second-order cone programming is rarely studied in the reported literature for the recent. Therefore, according to computer vision theory, spherical geometry, optimization theory and error theory, the proposed project can be divided into three main tasks described as fellows..1) Spherical geometry invariants describing the structure of objects are constructed from the corresponding figures of circle and quadrilaterals on the fisheye image and on the unit spherical image. .2) Fisheye camera pose estimation model is set up and is solved according to spherical geometry invariants and second-order cone programming..3) Source and property on fisheye camera pose estimation error of the proposed algorithm are analyzed and discussed under the different projective model. Image feature points which are less effect on errors are selected as input, or error corrected strategies are proposed to improve the accuracy of fisheye camera pose estimation.

随着传感器和网络技术的迅速发展及数字图像设备的普及，基于视觉的定位技术已经在机器人导航、虚拟现实、视觉监控等领域得到了广泛应用。近些年来，针孔摄像机由于视野范围较窄，在根据连续图像对其定位时关键特征难以持续捕捉，从而影响定位准确性；而具有大视角的全向摄像机定位吸引了许多研究人员的关注，但文献中还很少有利用球面几何不变量及二次锥规划优化研究鱼眼摄像机定位问题。因此，应用计算机视觉理论、球面几何学、误差理论、最优化理论等，本项目拟：1)根据几何图形如圆、四边形等在鱼眼图像及其单位球面投影间对应，构建易于描述物体结构特征的球面几何不变量；2)基于球面几何不变量及二次锥规划的鱼眼摄像机定位优化模型的建立与求解；3)探讨分析给出的定位算法在不同模型下的定位误差分布及在同一模型下的误差传递规律，并根据定位误差分布规律选择对定位误差影响小的输入图像点、或提出定位误差修正策略，以提高鱼眼摄像机的定位精度。

在计算机视觉领域，机器人定位与制图一直是热点研究问题之一，而摄像机定位、三维重建等其中的关键问题。本项目主要研究了：(1)基于鱼眼图像中的几何信息，估计出鱼眼摄像机的参数、空间三维位置，并引入二次锥规划方法进行了参数优化；该方法根据几何不变量能够较好地估计出摄像机的位置信息。(2)通过应用多幅图像间对应关系，提出了一种基于四个点顺序保持约束及相似性测量的单应矩阵估计方法。(3)假设图像上特征点提取误差服从椭圆分布的情况下下，提出了一种尺度自适应椭圆核一致性估计子和归一化椭圆权重的EWL-M优化算法，该方法能够从图像经检测的特征点中筛选出更号的内点，以估计出更精确的单应矩阵模型。(4)针对二维图像检测在缺失深度距离信息方面的不足，提出了一种基于深度信息的钢板表面三维检测方法。(5)提出了一种基于共面圆的距离传感器与相机的组合系统相对位姿估计方法，该方法不需要进行特征点的匹配，利用射影不变性来获取相机与三维距离传感器的位姿。本项目的部分研究成果在移动机器人、自动驾驶或无人机等领域中有着潜在的应用前景。