【作者】 李迪；

【导师】 续志军；

【作者基本信息】 中国科学院研究生院（长春光学精密机械与物理研究所） ， 机械电子工程， 2012， 博士

【摘要】 从机械稳像技术、光学稳像技术发展到电子稳像，电子稳像技术以其集成度好、能量消耗低、控制复杂度相对简单、稳像精度高、成本小、应用范围广、灵活便携等众多优势已成为稳像技术未来的发展趋势。本课题来源于中国科学院知识创新重要工程项目，对微型飞行器的电子稳像技术进行了深入研究，提出了解决微型飞行器图像序列抖动的有效方法。本文在研究原有电子稳像算法的基本原理、主要流程、核心环节(运动估计)的基础之上，结合本文旋翼式微型飞行器自身的特征限制和实际应用中的机载环境，创新的提出并实现了从运动估计和运动预测两个角度分别采用基于单应性透视投影模型结合SIFT算子和基于自组织递归区间二型模糊神经网络两种解决方案对机载抖动图像序列进行稳像处理，并通过实际的实验和系统的仿真分别对两种方案进行了有效性验证。在基于单应性透视投影模型结合SIFT算子的稳像方案中，本文分析了传统的基于仿射变换的图像运动模型不适用于本文旋翼式微型飞行器的原因，基于单应性透视投影原则,提出并推导出本文的图像运动模型。此模型包含丰富的图像运动信息，完全满足本文稳像系统的要求。在局部估计环节，应用了SIFT算子。通过实验验证，SIFT算子对具有一定几何形变、场景复杂度较高的图像局部特征都能进行准确的匹配，并在一定程度上能排除把小运动物体作为特征点定位。在基于自组织递归区间二型模糊神经网络的稳像方案中，本文创新的提出了运动预测的方案。传统的电子稳像技术是基于运动估计的方法，其发生在拍摄之后，与运动预测的最大区别就在于运动预测发生在拍摄之前。由于本文机载成像设备受风力和自身马达影响最大，因此，抖动具有一定的规律性。本文采用模糊神经网络的函数逼近及学习能力模拟出抖动规律，预测出在未来时刻机载成像设备抖动的位置，对其进行补偿，从而达到稳像的目的。本文对稳像系统可能的实现方案进行了对比分析，最终采用ARM基站+FPGA芯片相结合的嵌入式方式实现实时数字图像稳定系统的总体方案。在本文的最后给出了旋翼式微型飞行器稳像算法效果评价，阐述了图像序列稳定质量的评价方法，用实拍的航摄图像序列进行了电子稳像处理，证明了嵌入式方案的实时性、稳定性和所提出的稳像算法的有效性。更多还原

【Abstract】 Electronic Digital Image Stabilization (EDIS) has become the future trend ofdevelopment, because of its advantages of good Integration, low energyconsumption, simple control complexity, and high image stabilizing precision,reasonable price, wide range of applications, flexible and portable. This thesis,supported by the directional project with knowledge innovation and importantengineering of Chinese Academy of Sciences, has put forward the effective methodfor realizing stabilization of jitter image sequences. Intensive study of the techniqueof electronic image stabilization of MAV has been carried out.This thesis elaborated the fundamental principle and main processes of EDIS,especially for the key steps (motion estimation) adopted by overseas electronicimage stabilization system. On this basis, combining with MAV‘s characteristics andairborne environment, airborne jitter of image sequence were processed by twosolutions which based on homography with sift algorithm and Fuzzy neural networkrespectively. At the end of the relevant sections, two solutions were verified throughthe actual experiment and the simulation of the system.The first solution analyzed the reason why the traditional image motion modelwasn’t suitable for MAV, put forward and deduced the image motion modelaccording to homography perspective projection principle. In the local motion estimation step, SIFT operator was verified that certain geometric deformable, scenecomplexity of higher image local characteristics can be accurately match, and get ridof small moving objects as a feature location.The second would focus on motion prediction. The traditional EDIS based onmotion estimation after the shooting, that is to say, to calculate the image sequencesbetween adjoining frames of the relative displacement, and then implementation ofcompensation. Since this article airborne imaging equipment jitter was caused bywind and its own motor, and therefore, the jitter has certain regularity. This solutionusing fuzzy neural network function approximation and learning ability to simulatethe jitter law predicted the location of jitter airborne imaging equipment in the nextmoment and then, compensated the difference.In accordance with the system requirements it has been designed that ahigh-performance embedded image stabilization system based on ARM and FPGA.For the purpose of evaluation, we propose several performance measures to describethe fidelity, speed, and range of displacements supported by such systems. Thesemeasures can also be used as development tools to determine the influence of certainmodules in the overall performance of a system.更多还原