【作者】 孙瑾；

【导师】 秦小麟； 顾宏斌；

【作者基本信息】 南京航空航天大学 ， 计算机应用技术， 2008， 博士

【摘要】 飞行模拟器是航空工业的必要装备,是飞行员训练、考核的基本地面训练设备。与传统模拟机相比,基于虚拟现实的轻型飞行模拟器体积小巧,价格低廉,具有巨大的应用潜力。本文围绕“轻型飞行模拟器原型系统”课题,为进一步完善轻型飞行模拟器系统以满足高端需要,对虚拟现实中,特别是虚拟视景构建和虚拟环境交互操作中的图像关键技术进行了研究。论文的主要研究内容包括:(1)不同于传统模拟机,轻型飞行模拟器采用头盔式视景显示设备,舱内视景也需要由计算机生成,同时戴上头盔后飞行员无法看到舱内实景,造成交互操作困难。根据飞行模拟器视景特点和交互要求,设计了轻型飞行模拟器虚拟现实环境系统的体系结构。其中,采用传统几何建模方法生成舱内视景,便于实现交互;利用基于图像绘制技术构建舱外视景,提高真实感。并提出半虚拟现实座舱方案进一步完善虚拟现实系统手部人机交互性能。(2)图像特征信息提取是虚拟现实环境构建过程中涉及的基本和重要的图像处理技术。通过对图像边缘特性的分析,本文选择边缘信息进行特征提取,以减少冗余,提高图像处理技术的实时性和鲁棒性。传统微分算法对噪声敏感,对此提出一种基于方向信息抗干扰性边缘检测算法,并通过实验证明该算法在定位精度和抗噪性方面优于传统方法。同时对边缘中常用的直线特征,针对Hough变换只能精确到像素级的不足,提出一种基于Hough变换直线检测改进算法,在保持较高抗噪性的同时,提高检测精度,为后续得到更加准确和鲁棒的图像处理结果提供基础。(3)摄像机标定结果对舱外视景构建和手部交互定位都有指导意义。本文通过对现有摄像机标定技术进行归纳总结,在传统标定方法基础上提出了两种摄像机标定新方法。其中,基于消隐点的摄像机标定改进方法只利用2个方向消隐点确定摄像机内参数和方位,提高了利用消隐点进行标定的实用性;基于矩形的标定方法由矩形的四个顶点的像构造无穷远平面上绝对二次曲线的像ω的约束方程标定摄像机内参数。两种标定方法完全摆脱了三维空间点与二维图像点的匹配,简化了标定过程,同时无需进行特征匹配计算图像和空间平面间的单应矩阵,直接利用消隐点和矩形顶点就可确切的求出旋转矩阵和平移量,不会丢失摄像机外参数,可以准确定位摄像机,对手指定位具有重要意义。模拟和真实图像实验结果证明了两种标定方法的正确性和可行性。(4)舱外全景图构建过程中,图像拼接是关键技术之一。图像拼接的主要问题是确定相邻图像间的对应关系,即图像匹配。考虑Hausdorff距离易于计算,容错率高,采用修正后的Hausdorff距离构造相似性测度,提出了一种基于方向信息的鲁棒型Hausdorff距离的匹配方法。实验表明,该方法加快了匹配过程,提高了抗噪性能,并能够准确匹配含有遮挡和伪边缘点的图像,从而解决了基于传统Hausdorff距离匹配方法因噪声点、伪边缘点和出格点而造成的误匹配问题。(5)现有虚拟现实交互设备价格昂贵,而且无法提供逼真的触觉等丰富感知。为实现自然的人机交互方式,提出半虚拟现实座舱方案,设计了基于摄像机标定原理的戒指式视频测量手指定位技术和基于计算机视觉的虚拟现实手指定位技术两种技术方案。手部跟踪定位的首要任务是从摄像机拍摄的视频序列中进行运动目标(手部)检测,为此提出一种基于梯度方向信息的运动目标检测算法。实验结果表明,该算法克服了传统帧差算法不能准确定位目标的缺点,在室内外背景下均能准确的提取完整的运动目标轮廓。更多还原

【Abstract】 Flight simulators are essential pilot training-and-checking equipments in aviation industry. Compared with traditional flight simulators, Virtual reality based light flight simulators, i.e. VR simulators, have great potentialities in the applications for compact structure and relatively low price. Along with the project of“Prototype of Light Flight Simulator System”, several key technologies of image processing for virtual reality, especially for virtual scene generation and human-machine interaction of the light flight simulator, are studied to meet the high-end demands. The contributions of this dissertation are mainly summarized as follows:Firstly, the light Flight Simulator is equipped with Head Mount Display (HMD) to display virtural scene. The pilot with HMD does not see the real environment inside the cockpit,which needs to be generated in VR scene. Moreover, it is difficult to interact between human and machine. According to the specialties of virtual scene and the demand for interaction in light flight simulator, the structure of light flight simulator’s virtual environment is constructed. GBMR and IBR are respectively used to generate the inside and outside scenes of the cockpit. The concept of semi-virtual reality cockpit is proposed to achieve a natural human-machine interaction.Secondly, the pre-processing method of extracting feature from image is the basis image processing included in VR. In order to improve the real-time and robustness and reduce redundancy for image processing, this dissertation extracts edge feature for image pre-processing. For avoiding being sensitive to noise as traditional edge detectors, a direction-based edge detection method is put forward. Experiments prove that it has the superiority in removing noise and detecting edge accurately. Furthermore, considering Hough transform line detection only can locate in the range of pixel, an improved method of line detection based on Hough transform is proposed to enhance the precision of line detection. Experiments show that it improves the detection accuracy while maintaining efficient capability of anti-noise. It is also the base and prerequisite for some key problems,e.g. camera calibration and image registration.Thirdly, camera calibration plays an important role in image registration and finger-tracking. Two new camera calibrations are provided in this disseration. One is vanishing point (VP)-based calibration, the other is rectangle-based calibration. VP-based calibration directly obtains all camera parameters only by two vanishing points. Compared with classical VP-based calibration, it enhances the practical application of VP-based calibration. In rectangle-based calibration, linear constrains on camera intrinsic parameters are set up via the absolute conic imageωand the images of rectangle vertexes. Compared with existing methods, the two proposed avoid complex image matching, which simplifies the calibration process and decreases the computational complexity. What is more important, the two calibrations get all camera extrinsic parameters without computing homography of planes between image and space. Extensive tests with real and synthetic images show that two proposed approaches are both accurate and robust.Fourthly, image stitching is one of the important techniques to the study on cylindrical panoramic image mosaic. The main problem of imagestitching is correspondence, that is image matching. Considering the low computational complexity and high fault tolerance of Hausdorff distance, a new image matching method using direction-based robust hausdorff distance (DRHD) is presented. DRHD constructs a similarity measure based on the proposed improved Hausdorff distance. The experimental results show that the proposed algorithm speeds up the matching process and it improves the resistance to noise. In addition, this method matches the image occlusions correctly and overcomes the mismatching problems that induced by noise, spurious edge segments and outlier points, which demonstrate that the proposed method is feasible and effective.Finally, existing interaction equipments are expensive and do not provide realistic and rich sense, such as touch and force. To satisfy natural interaction in VR system, semi-virtual reality cockpit is put forward. In addition, two finger-tracking techniques are provided. One is the ring type video measure base on camera calibration; the other is finger tracking based on computer vision. For the demand of real-time in finger-tracking system, a gradient direction based detection algorithm is provided to detect moving object (hand) from a video image sequence,. Experimental results show that the proposed algorithm overcomes the shortcoming of not correctly detecting moving object of traditional frame difference algorithms and can effectively and accurately extract moving object contour from indoor and outdoor environments更多还原