【作者】 关柏青；

【导师】 王石刚；

【作者基本信息】 上海交通大学 ， 机械设计及理论， 2007， 博士

【摘要】 运动及振动参数的机器视觉测量方法广泛应用于制造业、自动化以及医疗领域。机电系统动态性能测试以及视觉伺服等工业应用领域中的目标运动速度及振动频率远远高于一般应用场合,因此高速运动估计和高频振动测量成为运动和振动参数测量的一个重要研究内容。针对高速运动估计和高频振动测量中的几个常见问题,如主动序列运动模糊图像采集和建模以及2D高速平移运动、2D高频面内振动、高速定轴转动和高频角振动以及高速离面运动的测量,本文开展了系统和深入的研究。1.高速运动或高频振动测量中,目标图像平面上将产生运动模糊现象。这种运动模糊一般被视为不利于图像分析的一种退化,而本文根据生物视觉中存在的“由模糊感知运动”的机制,认为图像模糊现象携带了重要的运动信息,可以在运动和振动参数测量中主动加以利用。同时,根据主动视觉原理,本文认为图像的运动模糊程度可以由视觉系统主动控制,并设计了一种主动序列运动模糊图像采集系统,该系统可以根据图像的运动模糊程度及饱和程度来自动预测和控制摄像机的曝光时间、采样周期和光电转换系数,从而产生携带足够运动模糊信息的图像或图像序列,同时避免图像的亮度饱和现象。此后,根据该图像采集系统的特点,本文提出了主动序列运动模糊图像的一般数学模型,并分析了目标作面内平移、旋转以及离面等不同形式运动时的运动模糊图像模型。2.高速平移运动是高速运动及高频振动中一种简单但十分重要的形式。由于运动模糊现象的存在,目前平移运动分析中常用的块匹配法难以直接应用于高速平移运动分析中。本文根据生物视觉中存在的“由模糊感知运动”的机制以及生物运动知觉中的时空整合机制,提出了一种基于主动运动模糊图像及几何矩的2D高速平移运动估计方法。该方法首先利用几何矩模拟生物运动知觉中的空间整合机制,并推导出表征平移运动模糊图像的几何矩以及运动函数之间关系的定理,根据该定理提取运动模糊图像中的运动信息,从而实现运动估计。本文的算法可简单的利用两帧运动模糊图像实现2D加速运动的估计,而现有方法需要至少三帧图像才能得到同样的结果。3.振动测量是机械系统测试特别是微机电系统测试中的重要内容。常用的振动测量方法是采用辅助的频闪光源采集目标的清晰图像序列并加以处理。而现有的基于运动模糊图像的方法仅针对振动周期远小于摄像机曝光时间的高频振动振幅测量方法进行了研究,并未涉及振动周期大于曝光时间的低频振动(以下及正文中简称为低频振动,但这里的低频并不是传统意义上的低频)测量问题。本文着重分析了高频及低频的振动模糊图像特点及其与非往复运动的不同之处,从而根据基于运动模糊及图像矩的高速平移运动估计方法分别设计出面内高频及低频振动测量的具体方法。本文的算法可以根据一幅清晰图像及一幅运动模糊图像或两幅连续的运动模糊图像计算出低频振动的振幅,频率,相位和振动方向以及高频振动的振幅及振动方向。同时,在图像噪声水平较高时,提出了采用短运动模糊图像序列信息融合方法来测量振动参数的思路,该方法可显著提高测量精度。此外,设计了一种简便的视觉系统参数控制方法,以提高振动频率未知时的自适应性。4.定轴转动及角振动普遍存在于许多工业应用领域。本文根据圆周运动模糊图像的直角坐标系模型和极坐标系模型,分别提出了两种基于主动序列运动模糊图像及几何矩的定轴转动及角振动参数的估计算法:其一是利用运动模糊图像的直角坐标系模型估计转动或高频角振动的参数。该方法首先推导出直角坐标系下运动模糊图像的几何矩与圆周运动函数的关系定理,然后根据该定理提取图像平面上的圆周运动模糊信息,从而计算出转动或角振动的参数。其二是直接将平移运动及振动估计方法映射到的极坐标平面上,从而实现转动及角振动的测量。与现有的定轴转动及角振动测量方法相比,本文提出的方法可测量曝光时间内存在较大角位移的圆周运动。5.离面运动测量的常用方法是激光多普勒仪、电子或数字散斑干涉以及光学剪切干涉,它们均需要额外的光源及干涉光路。通过分析离面运动模糊的成因,本文提出了散焦模糊与图像放缩运动模糊相复合的离面运动模糊图像模型,并推导出离面运动模糊图像的几何矩以及离面运动函数之间的关系定理,从而根据该定理提取离面运动模糊图像中的离面运动特征并实现离面运动估计。本文提出的方法可以较准确的测量离面运动的参数。更多还原

【Abstract】 Vision-based measurements of motion and vibration are widely used in manufacturing, automation and medical field. In several industrial fields such as dynamic behavior testing of electromechanical system and visual servo, the velocity of the motion or the vibrational frequency are usually higher than those in ordinary fields. Therefore, measurements of high-speed motion and high-frequency vibration become important aspects for motion analysis and vibration measurement. This paper is focused on some basic issues in high-speed motion analysis and high-frequency vibration, such as the acquisition and the modeling of actively blurred image sequence, 2D high-speed translational motion analysis, in-plane translational vibration measurement, high-speed rotation and angular vibration measurement and high-speed out-of-plane motion estimation, and so on, for deep research.1. Images tend to be blurred when the motion velocity or the vibrational frequency of the object is large. The so-called motion blur is usually treated as image degradation. But from the“motion-from-blur”mechanism existing in biological vision, we infer that motion blur is an important cue for motion analysis and can be actively used in motion and vibration measurements. Furthermore, according to the theory of active vision, this paper considered that the degree of the motion blur can be controlled by the vision system, and designed an acquisition system of actively blurred image sequence. This image acquisition system can predict and control the exposure time and sampling period according to the degree of motion blur and saturation of previous images. Hence, it can acquire image sequences carrying sufficient motion blur information without image saturation. Considering the characteristics of the image acquisition system, we proposed the general model for the actively blurred image sequence, and analyzed variant models for different motions such as in-plane translation and rotation and out-of-plane motion.2. High-speed translational motion is a simple but very important form among high-speed motion and high-frequency vibration. Block matching, which is widely used in translational motion analysis breaks down when the high velocity causes motion blur in image. According to the“motion-from-blur”and spatio-temporal integration mechanisms existing in biological motion perception, this paper presented a method for estimating 2D high-speed translational motion based on actively blurred images and geometric moments. This approach relies on geometric moments to achieve spatial integration, derives a theorem for the relationship between the geometric moments of the motion blurred image and the translational motion, and extracts the motion blur information of the images to realize motion analysis. The proposed algorithm can calculate the velocity and acceleration of the motion from only 2 successive frames of motion blurred images while existing approaches estimate an accelerated motion from at least 3 images.3. Vibration measurement is very important for testing of mechanical system, especially for the testing of micromechanical system. General techniques employ stroboscopic illumination to“freeze”the vibrating object and acquire unblurred images for vibration measurement. And existing approaches based on motion blurred images only dealt with high-frequency vibration which period was much shorter than the exposure time, and did not mention the low-frequency cases which period was longer than the exposure time (Here note that the low-frequency does not really mean the vibrational frequency is low). This paper laid emphasis on analyzing the difference between the blurred images with low- and high-frequency vibrations and non-vibrational translation, and derived the algorithms for measuring in-plane vibration with low- and high-frequencies, respectively, based on the high-speed non-vibrational translation analysis method mentioned above. This method requires only one motion blurred image and an unblurred image or two successive frames of blurred images to calculate the parameters of the low-frequency vibration as well as the amplitude and direction of the high-frequency vibration. With the existence of high-level noise, we employed data fusion on a short sequence of motion blurred images to improve accuracy. Further, this paper designed a method for controlling the setting of the vision system to improve the adaptability of this algorithm when the frequency was unknown.4. Monoaxial rotation and angular vibration exist in many industrial applications. This paper proposed two algorithms based on motion blurred image sequence and geometric moments for measuring monoaxial rotation and angular vibration, according to the Cartesian coordinate system and polar coordinates, respectively. The one is based on the Cartesian model of the motion blurred image. It derives the theorem for the relationship between the geometric moments of the motion blurred images with Cartesian coordinates and the circular motion, then extracts the motion blur information on the image plane, and calculates the parameters of monoaxial rotation and angular vibration. The other maps the algorithms for measuring translational motion and vibration to the polar plane to realize measurements of the monoaxial rotation and the angular vibration. Compared with existing techniques for measuring monoaxial rotation the angular vibration, the proposed algorithms can measure circular motion with large angular displacement within the exposure time.5. The most widely used techniques for out-of-plane motion measurement are Laser Doppler Velocimetry (LDV), Electronic Speckle Pattern Interferometry (ESPI) and Shearography. They all need additional laser or other light source and interferential path. By analyzing the image blur caused by out-of-plane motion, this paper proposed the motion blurred image model for out-of-plane motion, which was combined with defocus blur and motion blur. We derived the theorem for the relationship between the out-of-plane motion blurred image and the out-of-plane, and then extracted the information of the out-of-plane motion to analyze of the out-of-plane motion. The proposed method can measure the parameters of the out-of-plane motion correctly.更多还原