【作者】 刘景正；

【导师】 余旭初；

【作者基本信息】 解放军信息工程大学 ， 摄影测量与遥感， 2011， 博士

【摘要】 面向基于动态环境的无人飞行器目标实时监视需求,以快速、准确获取运动目标的位置和场景属性为目的,本文针对可见光、红外序列影像的特点,从无人飞行器影像的不变性匹配、仿射变形目标匹配关联、跨场景影像配准、运动目标检测与跟踪等方面做了较为深入的研究。研究的内容和创新点主要有:1,分析了无人飞行器目标监视技术的组成和流程,设计了目标监视系统架构。2,分析了局部不变性稀疏特征匹配各方法的原理,针对无人飞行器序列影像的应用特点,提出空间分布控制方法来设置SIFT匹配的取舍参数,使匹配点更加均匀;在搜索策略上采用改进的k-d树提高处理速度;并以实验对比了各方法优缺点。3,引入了基于ASIFT经纬度模拟的大倾角影像地物目标关联算法。在低分辨率层模拟目标影像成像时的倾斜纬度和旋转经度,使用尺度不变算法获得与搜索影像相似的模拟影像;完成选定模拟影像在高分辨率层对应影像的精确匹配。克服了无人飞行器大倾角拍摄时带来的仿射变形影响,保证了序列影像目标框定的稳定。4,提出一种基于稠密SIFT流的跨场景影像配准算法。将稀疏特征改进到逐个像素对应的稠密特征,同时保持空间离散性,并将多尺度信息量度量引入到二维影像中,由粗到精完成不同场景的影像对应关系。影像不同场景配准实验证明,本方法完成了传统像素级影像配准所不能完成的任务。5,提出结合空间结构与MeanShift的目标跟踪算法。针对红外影像特点,采用了小波滤噪的方法降噪。对序列影像相邻帧采用SURF配准的方法检测运动目标,依据Kalman滤波方法对运动目标预测;从搜索区域提取中心点空间结构描述符,与颜色直方图连接,结合MeanShift修正跟踪目标位置。解决了目标跟踪时的遮挡、大小变化等问题。6,研究了粒子滤波算法的原理和过程,结合Adaboost和混合粒子滤波完成无人飞行器运动目标的检测与跟踪。利用学习型Adaboost建议分布快速检测到场景中的目标,生成运动目标检测结果。采用混合粒子滤波的方法,用Adaboost生成的假设检验和目标动态模型信息组成混合模型来构建建议分布函数,实现运动目标的实时跟踪。更多还原

【Abstract】 Basing on the exigent requirements of the real-time object surveillance of dynamic environment for unmanned aerial vehicle(UAV), and aiming at applications of object location and the scene property acquired by UAV, this dissertation makes a study on several key techniques for optical and infrared sequential imagery, such as invariability matching, object orientation of affine distortion, image registration across scenes, motion object detection and tracking,etc. The works achieved in this dissertation are mainly as follow:1,The components and the process of object surveillance technique for UAV is analyzed, with the framework designed too.2,The theoretics of each matching method of sparse local invariant features is analyzed. Aiming at characteristic of applications for UAV sequential imagery, the space distribution controlling method is proposed to adjust the parameters of SIFT, which makes the matching space more uniform. The improved k-d tree method is adopted to accelerate processing. Each matching method is implemented, and the advantage and disadvantage of each method are analyzed with the experimental results.3,A method of simulating longitude and latitude based on ASIFT is proposed, in order to find the orientation of object in oblique image. The procedure selects the affine transforms that yielded matches in the low-resolution process, then simulates the selected affine transforms on the original query and search images, and finally compares the simulated images by SIFT. The effect of affine distortion by great tilt of camera axes on UAV can be overcome, and the stabilization of object orientation of sequence imagery can be guaranteed, too.4,SIFT flow is introduced into the method for registering an image to its nearest neighbors in a large image corpus containing a variety of scenes. The SIFT flow algorithm consists of matching densely sampled, pixel-wise SIFT features between two images, while preserving spatial discontinuities, accomplishing the correspondence densely across scenes of coarse-to-fine. Experiments of registration across scenes of UAV sequential imagery prove that this method can complete assignments which the tradition method is not able to.5,A method that combines center point featrue descriptor and MeanShiftis is proposed. According to the noise characteristics of infrared image, the wavelet method is used to reduce the noise of infrared image. After registering with SURF matching method, the motion object can be predicted by Kalman filtering. A new histogram can be made by combining the color histogram and center descriptor extracted in search area, then with MeanShift modifying, object tracking can be achieved. Also, the problem of object blocked and size changed is solved in this dissertation.6,The principal of Particle Filtering is studied. An approach of combining two well-developed algorithms: mixture particle filters and Adaboost, is adopted. The learned Adaboost proposal distribution allows us to quickly detect object, while the filtering process enables us to keep track of the motion object. We construct the proposal distribution using a mixture model that incorporates information from the dynamic models of object and the detection hypotheses generated by Adaboost. The result of interleaving Adaboost with mixture particle filters is a simple, yet powerful and fully automatic multiple object tracking system.更多还原