基于宽视场拼接成像的目标分割与跟踪算法研究

发布时间：2018-06-23 01:39

本文选题：图像拼接 + 统一计算设备架构　；参考：《长春理工大学》2016年博士论文

【摘要】：无论是民用领域的矿产资源勘查、土地规划利用、环境监测、海洋开发、气象预报及地理信息服务还是军事领域侦察监视、精确制导、超视距攻防对抗等都需要有足够宽的视场和足够高的分辨率以完成对目标的广域范围监测、搜索和跟踪。对基于宽视场高分辨成像系统海量图像数据的处理、分析和利用是该类系统建构的核心价值所在。其中,高分辨率传感器拼接成像过程中将涉及到对海量数据的实时处理,计算过程有着极高的复杂度,因此,图像拼接算法的准确性和实时性将成为影响系统性能的主要因素之一。此外,对于宽视场高分辨场景下动态目标的跟踪技术也成为后期图像分析的研究热点。同时,由于这类系统应用的环境自身的复杂性(背景变化、光照变化、阴影变化等)和动态目标固有的一些特性(非刚体、姿态多变等),使得可实用的目标跟踪技术仍然非常具有挑战性。针对以上需求,本文围绕宽视场成像系统的图像拼接和目标跟踪问题开展研究,主要研究工作如下:本文采用了一种将先验信息和统一计算设备架构(CUDA)相结合的自适应并行加速算法用于提高大视场全景拼接成像的实时性。在图像拼接之前,先利用高精度标定平台对各成像单元的重叠区域进行预标定。标定之后,利用基于CUDA的快速鲁棒特征检测方法提取参考图像与待配准图像的候选特征点集,再利用基于随机KD-Tree索引的近似最近邻搜索算法选取参考图像与待配准图像的初始匹配点对,本文还采用了基本线性代数运算子程序用于加速算法搜索速度。对于参考图像与待配准图像误匹配点对的删除和空间变换矩阵的参数估计,本文采用的是一种在传统的渐近式抽样一致性算法基础上改进的基于CUDA的并行算法。实验结果表明本文采用的算法极大地提高了图像拼接速度,可以满足图像拼接实时性的工程应用要求。为了对场景中动态飞行目标进行识别,提出一种基于混沌双种群进化策略的图像分割方法。利用进化策略能从选定的初始解出发,通过逐代迭代进化逐步改进当前解,直至最后收敛于最优解或满意解的特点和优势,将其用于图像分割阈值最优解的求解上。为了克服传统基于阈值的图像分割方法的缺点,例如较高的复杂度和早熟问题,本文提出了一个高效的基于进化策略的图像分割算法,它通过使用多种群进化策略来计算阈值。在进化过程中同时存在局部种群和全局种群两个群体,进而确保算法的全局和局部搜索能力。该算法的每一步迭代过程中,首先,基于混沌理论生成若干个初始个体,并将这些个体分别加入局部种群和全局种群,计算这些个体的适应度函数值。然后,将选择、重组、变异等进化操作算子作用于局部种群和全局种群,进行迭代进化,进化后的个体集合中选择最好的若干个体放入局部种群,其余放入全局种群,直至收敛。最后,种群中的最优个体即为所求的解。实验结果表明,本文提出的方法比传统的遗传算法有着更快的收敛速度。种群多样性信息能有效指导进化策略的进化过程,因此本文又提出了改进的混沌双种群进化策略算法,采用了多动机强化学习算法设定初始种群和本地种群数值,动态学习种群比例,以使进化策略的局部搜索能力和全局搜索能力进一步均衡化。动机层的引入为先验知识和领域知识的引入提供了条件,由此可以加速强化学习的学习进程。本文根据图像分割问题实际,定义了动机集合,采用了MMQ投票(MMQ-voting)方法用于指导智能体动作的选择策略。经过实验验证,本文采用的多动机强化学习方法能使强化学习以较快的速度收敛于最优动作策略,从而使种群个体多样性保持在一个合适的状态,有助于进一步提高图像最优分割阈值的搜索效率。为了对场景中动态飞行目标进行跟踪,提出一种基于强化学习的动态目标跟踪方法,将目标跟踪问题建模成强化学习问题,并提出了一个两阶段强化学习算法用于图像中的目标跟踪。我们设置了多个追踪智能体来跟踪图像中的目标,在算法的每一步中,首先对每个追踪智能体进行动态子任务分配,即先是给每个追踪智能体动态分配一个子目标,之后每个追踪智能体根据其当前的子目标选择其行动。学习算法将学习过程划分为两个部分,一个是学习任务分配的策略,另一个是学习动作选择的策略,每个追踪智能体通过共享Q函数来共享所学知识、提高学习效率。实验结果验证了该方法的有效性。
[Abstract]:Whether it is the mineral exploration in the civil field, land planning and utilization, environmental monitoring, marine development, weather forecast and geographic information service or military field surveillance and surveillance, precision guidance, and over the horizon attack and defense confrontation need to have wide field of view and high enough resolution to complete the wide range monitoring, search and tracking of the target. The analysis and utilization of massive image data based on wide field of view high-resolution imaging system is the core value of this kind of system construction. Among them, high resolution sensor splicing will involve real-time processing of mass data and high complexity in the calculation process. Therefore, the accuracy and reality of the image mosaic algorithm will be true. Time character will be one of the main factors that affect the performance of the system. In addition, the tracking technology for dynamic targets in the wide field of view high resolution scene has also become a hot topic in the later image analysis. At the same time, due to the complexity of the environment itself (background change, illumination change, shadow change, etc.) and some inherent characteristics of the dynamic target, the application of this kind of system The practical target tracking technology is still very challenging. Aiming at the above requirements, this paper focuses on the problem of image mosaic and target tracking in the wide field of view imaging system. The main research work is as follows: a combination of prior information and unified computing device architecture (CUDA) is used in this paper. The adaptive parallel acceleration algorithm is used to improve the real-time performance of the panoramic mosaic imaging in large field. Before the image splicing, the high precision calibration platform is used to pre calibrate the overlapped regions of the imaging units. After calibration, the CUDA based fast robust feature detection method is used to extract the candidate feature points of the reference image and the image to be registered. Set, and then use the approximate nearest neighbor search algorithm based on random KD-Tree index to select the initial matching points of the reference image and the image to be registered. This paper also uses the basic linear algebra operation subroutine to speed up the search speed. It is estimated that this paper uses an improved CUDA based parallel algorithm based on the traditional asymptotic sampling consistency algorithm. The experimental results show that the algorithm used in this paper can greatly improve the speed of image stitching, and can meet the engineering application requirements of the real-time image splicing. An image segmentation method based on chaotic dual population evolution strategy is proposed. Using the evolutionary strategy, it can gradually improve the current solution by iterative evolution from the selected initial solution, and finally converge to the characteristics and advantages of the optimal solution or satisfactory solution, and apply it to the solution of the threshold optimal solution of the image segmentation. The shortcoming of threshold image segmentation methods, such as high complexity and precocious problem, presents an efficient image segmentation algorithm based on evolutionary strategy. It uses a variety of cluster evolution strategies to calculate threshold. In the process of evolution, there are two groups of local population and whole local population, which can ensure the overall situation of the algorithm. In each iteration of the algorithm, first of all, a number of initial individuals are generated based on the chaos theory, and these individuals are added to the local population and the global population to calculate the fitness function values of these individuals. Then, the evolutionary operators such as selection, reorganization and mutation are used to act on the local population and the global population. The optimal individual is placed in the local population and the rest is placed in the global population to converge. Finally, the optimal individual in the population is the solution. The experimental results show that the proposed method has a faster convergence rate than the traditional genetic algorithm. The population diversity information can be effective. In order to guide the evolutionary process of evolutionary strategy, this paper also proposes an improved chaotic dual population evolution strategy algorithm, which uses a multi motivation reinforcement learning algorithm to set the initial population and local population values and dynamically learn the population proportion, so that the local search ability and the overall search ability of the evolutionary strategy are further balanced. The introduction of motivation layer is introduced. It provides the conditions for the introduction of prior knowledge and domain knowledge, which can accelerate the learning process of intensive learning. Based on the reality of the image segmentation, this paper defines the set of motivation and uses the MMQ voting (MMQ-voting) method to guide the selection strategy of the action of the agent. The method can make the reinforcement learning converge to the optimal action strategy at a faster speed, so that the individual diversity of the population is kept in a suitable state and helps to further improve the search efficiency of the optimal segmentation threshold of the image. In order to track the dynamic flight targets in the scene, a dynamic target tracking method based on reinforcement learning is proposed. The target tracking problem is modeled as a reinforcement learning problem, and a two stage reinforcement learning algorithm is proposed for target tracking in the image. We set up multiple tracking agents to track the target in the image. In each step of the algorithm, we first assign each tracking agent into the action state sub task, that is, each tracking intelligence first is given to each tracking intelligence. The learning algorithm divides the learning process into two parts, one is the strategy of learning task allocation and the other is the strategy of learning action selection. Each tracking intelligent body shares the learned knowledge by sharing the Q function, and improves the learning process. Learning efficiency. Experimental results verify the effectiveness of the method.
【学位授予单位】：长春理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】