【作者】 孙春凤；

【导师】 袁峰；

【作者基本信息】 哈尔滨工业大学 ， 仪器科学与技术， 2011， 博士

【摘要】 高速图像序列的运动目标检测是计算机视觉领域中富有挑战性的课题之一,在军事、航空航天和工业生产等领域有着广阔的应用前景,对其展开研究具有重要的理论价值和实践意义。但是,迄今为止,高速图像序列的运动目标检测技术研究仍然存在许多问题没有得到良好的解决,特别是随着应用要求的不断提高以及成像技术的飞速发展,图像数据量明显增加,如何解决好实时性和目标检测精度的矛盾,是高速图像序列运动目标检测技术研究的一个关键问题。课题“基于并行处理的高速图像序列运动目标检测技术研究”的目的就是以实时运动目标检测算法和并行系统结构设计为中心,全面深入地研究高速图像序列运动目标检测系统涉及的若干关键问题。在查阅大量国内外有关文献的基础上,论文概述了国内外运动目标检测方法的研究概况,包括时域运动目标检测方法、空域运动目标检测方法以及时空联合运动目标检测方法的研究现状,并分析了不同运动目标检测方法的特点。同时,论文对图像并行处理技术、图像传感器运动模型以及运动目标检测系统结构设计的国内外研究现状进行了综述,分析了不同系统结构的特点及适用场合。本课题在现有的运动目标检测技术理论研究的基础上,结合高速图像序列的数据特点,构建了并行的高速图像序列运动目标检测系统的体系结构。考虑到高速图像序列运动目标检测系统采集到的海量图像数据,不可避免地会受到各种噪声的干扰以及光照条件的变化对运动目标检测结果的影响,论文将部分重叠的思想引入图像预处理过程,提出了基于整数小波去噪和直方图均衡的部分重叠图像预处理(WHI)算法。该算法通过在整数小波变换后的高频子带进行逐点BayesShrink阈值去噪和中值滤波,降低高斯和椒盐噪声对图像质量的影响;同时,通过低频子带的亮度保持直方图均衡,解决了光照不均引起的图像质量下降问题。对于图像分块处理所带来的块效应现象,通过双线性插值加以消除。实验证明,论文提出的算法相比于传统的图像预处理算法,不仅具有较高的峰值信噪比(PSNR)值,而且算法的运算时间更短,占用的存储资源更少,更有利于硬件实现。高速图像序列采集过程中,图像传感器通常需要随着运动目标一起运动,因此图像传感器自身的扰动将直接影响运动目标检测结果的准确性。为了消除这些图像序列中的全局运动干扰,论文引入了匹配块预判的思想,提出了一种基于并行起点预测的三参数模型背景运动补偿(PCQHBS)算法。该算法首先利用梯度信息减少了“不可靠”宏块对于运动矢量估计的影响,然后将并行起点预测、搜索中止准则和自适应搜索模式确定方法融入基本的交叉六边形搜索算法估计运动矢量参数,并根据三参数运动模型进行背景全局运动补偿。实验证明,该算法相比于传统的背景运动补偿算法,不仅具有较高的峰值信噪比(PSNR)值,而且减少了算法的运算量,提高了算法的运算速度。利用对称差分和投影技术进一步得到图像序列中的运动目标区域。为了在海量图像数据中得到准确的运动目标外轮廓信息,论文在传统Snake模型的基础上,引入了轮廓点标注的概念,提出了一种基于标注的部分最优Snake模型轮廓检测(LP-Snake)算法。该算法增加了两个外部能量函数:边缘能量和向心能量,并将轮廓检测过程分为初始期、成长期和逼近期三个阶段。在运动目标外轮廓检测的三个阶段通过采用不同的能量函数和轮廓重心确定方法减少了算法的运算量。实验证明,该算法在确定运动目标外轮廓的过程中减少了实际参与计算的轮廓点数,提高了运动目标外轮廓提取速度。为了高效快速地检测高速图像序列中的运动目标,进一步减少系统占用的存储资源,论文根据构建的并行体系结构,将高速图像序列的运动目标检测任务多路并行地分配到FPGA或DSP中实现。课题所涉及的各子系统关键模块的硬件实验证明,该并行结构大大提高了系统的运算速度,并有效地减少了系统占用的存储资源。验证结果证明了论文设计的高速图像序列运动目标检测系统的可行性和有效性。更多还原

【Abstract】 Moving objects detection in high-speed image sequences is one of the most challenging techniques in computer vision, which is widely used in military, aerospace and industry. For the great value in theory and practice, various new theories and algorithms have emerged continously. So far, there are still many dificulities in practice. Specially, with the improvement of application requirements, as well as the rapid development of imaging technology, the system data are increased largely. How to solve the contradiction between real-time and precision is a key problem of moving objects detection in high-speed image sequences. The objective of this thesis is to research deeply some key problems for moving objects detection in high-speed image sequences focusing on real-time moving objects detection algorithms and parallel system structure design.After consulting a large number of domestic and foreign related literatures and references, this thesis provided an overview of moving objects detection method. Time domain, airspace domain and time-space moving objects detection method is described, and the characteristics of different methods are analyzed. Furthermore, image parallel process techniques, image sensor motion models and the current structural design of moving objects detection system are reviewed, and the characteristics and applied occasionsof the different structures were analyzed. Based on the existing moving objects detection technique theoretics, with the data characters of high-speed image sequences, the parallel system structure of high-speed image sequences moving objects detection system are designed.In view of the influence of various noises and the brightness change for mass acquired image data of the system, the partially overlapped idea is used in the process of image pre-process. An overlapped image pre-processing algorithm with integer wavelet transform and histogram equalization (WHI) is presented. Each high frequency transformed by integer wavelet is denoised by point-to-point BayesShrink thresholding and median filter to reduce the fluence of Gaussian noise, salt and pepper noise to image quality. Furthermore, brightness preserving histogram equalization is used in low frequency band to solve the problem of uneven brightness. Because of the overlapped image preprocessing algorithm, sub-images are needed to be interpolated by bilinear interpolation to solve blocking effects. Testing results proved that the algorithm have higher Peak Signal to Noise Ratio (PSNR) than tradition image pre-processing algorithms. The less time and the fewer memory space are needed. It avails the hardware realization of WHI algorithm. The image sensor needs to move with moving objects during the course of the high-speed image sequence data acquision. The perturbation motion of image sensor effects directly the accuracy of moving objects detection results. In order to avoid the global motion in image sequence, the matching block pre-judgment is used. The three parameters model background motion compensation algorithm with parallel initial position predict (PCQHBS) is presented. Above all, a gradient-based block prediction strategy is used to reduce the block number participated in matching. Then, the parallel prediction of initial search point, half-stop criteria and adaptive search modes is used in rudimentary cross-hexagon-searching algorithm to estimate motion vectors. The background motion compensation with three parameters model is used. Experiments proved that the algorithm have higher PSNR and speed than tradition background motion compensiton algorithms. Moving object areas can be chosen in the compensated image sequences by symmetric difference and projection technique.In order to acquire the accurate moving objects contour informations, the idea of labeled contour points is adopted, based on the traditional Snake model. A part-optimimum Snake model contour detection (LP-Snake) algorithm with labeled contour points is represented. The two outer energy functions are added into traditional Snake model: edge energy and centripetal energy. It can be devided into three phases: initial phases, building phases and approximation phases. The different energy functions and contour orthocenter determination method are used in three phases of contour detection. Experiment proved that the actual computed contour points are reduced, and the contour detection speed can be accelated.In order to detect moving objects in high-speed image sequences effectively and reduce the memory space of the system, the task is assigned into several FPGA or DSP according to parallel system structure. Hardware experiment of the key modules of each subsystem proved that the operating speed of the system is increased, and the memory space of the system is reduced effectively. The results validate the correctness and effectiveness of high-speed image sequences moving objects detection system.更多还原