图像视频中运动估计和分析
发布时间：2018-11-16 18:16
【摘要】：随着多媒体技术和传感器技术的飞速发展,图像和视频作为记录视觉的载体,对人们的生活、生产等产生越来越重要的影响。近十年来,在各个方面的记录视觉信息的需求,产生了爆炸式的图像和视频。在这些数据中,运动信息在图像和视频的生成和记录过程中都扮演着重要的角色,可以说覆盖了图像和视频处理的各个领域,而在语义上也涵盖了从低层的运动轨迹估计到高层的运动理解。具体来说,在图像生成过程中,相机运动会带来图像的模糊,而使得图像的信息熵增大,为了能够恢复出清楚的尖锐的图像,我们往往需要精确估计相机的运动。在图像的记录过程中,有可能因为物体或人的快速运动造成了图像内容的运动模糊。在记录过程中产生的运动模糊往往比清楚图像提供的信息更多因为其捕捉到了动态的物体叠加。视频通常可以看作是图像在时序上的叠加,类似地,在视频产生过程中,相机的运动往往会导致画面的晃动和模糊,无论在图像质量上和视频可观赏性上都有一定的影响。而在视频内容记录中,运动信息往往是视频存在的理由,分析其运动往往更关注高层的语义。本课题以运动为核心,对图像和视频中的运动估计和分析展开了深入研究,具体包括:图像生成过程中相机运动的建模和表征,模糊图像中的相机运动估计和图像复原,视频内容记录中的运动的多层次表征学习以及运动的快速分析。本论文的主要工作和创新点可以总结为以下几点:1.在图像成像过程中对相机运动的深层分析,我们提出了对运动核进行分解并独自优化的模型。该分解模式能够揭示相机成像的内在特性,从而以一种全新的角度观测经典的图像去模糊问题。为了展示该表征的优势,我们提出了轨迹随机扰动算法来优化运动核。在很多例子当中,我们发现现有的去模糊算法落入局部极值的时候,我们的算法通过独立优化相机轨迹能够取得较好的去模糊效果以及正确的模糊核。2.在图像成像过程中利用高亮区域能够较准确地刻画相机运动这一特点,我们针对夜景这一非常具有挑战性的场景结合高亮区域把该问题变得可行。我们提出了一个全新的框架有机地把从高亮区域中推断出的运动核和非高亮区域结合求解更准确的运动核,除此之外,我们提出了一个全新的函数化运动核表征从而较准确地从高亮区域推断运动核,我们提出了一个新的能量最小化方程能够自动地把提取的运动核分配给不同的区域以便进行非均匀去模糊。3.在视频内容记录中,我们侧重分析了视频内容中的重要的运动信息:摔倒动作检测,为了适应视频流场景下实时的运动分析,即实时的摔倒检测,我们把视频内容中运动信息按照"难易程度"分层,通过级联的方式进行动作检测,不同于传统的级联框架,该级联框架能够支持不同复杂度的特征。通过这种混合特征的级联框架,我们的系统在精确度和效率上能够达到较好的折中。除此之外,我们精细地设计了我们采用的特征,支持特征复用以及增量式更新从而能够对视频流场景具有较好地支持。最后,在摔倒动作检测的基础之上,我们进行了拓展从而能够支持一般种类的动作检测以及引入了更多种类的特征从而在精确度上有一个更好的提升。本文针对图像生成过程中和视频内容记录中的运动进行了深入的分析。大量的实验结果表明了我们对相机运动建模的有效性以及对视频内容中运动分层而快速检测的高效性。
[Abstract]:With the rapid development of multimedia technology and sensor technology, image and video, as the carrier of recording vision, have become more and more important to people's life, production and so on. In the last decade, the demand for recording visual information in various aspects has resulted in an explosive image and video. In these data, motion information plays an important role in both the generation and recording of images and videos, and can be said to cover various fields of image and video processing, and in the semanteme it also covers the motion understanding from the low-level motion track to the high-level motion. In particular, in the image generation process, the camera motion causes the blurring of the image, and the information entropy of the image is increased, and in order to be able to recover a clear sharp image, it is often necessary to accurately estimate the motion of the camera. during the recording of the image it is possible that the motion blur of the image content is caused by the rapid movement of the object or person. The motion blur generated in the recording process tends to be more information than the information provided by the clear image because it captures the dynamic object superposition. The video is generally considered to be a superposition of the images in time series, similarly, during the production of the video, the motion of the camera often results in the shaking and blurring of the picture, whether in the image quality and in the video. In the video content record, the motion information is often the reason of the existence of the video, and the analysis of its motion tends to pay more attention to the semantics of the higher layer. This paper studies the motion estimation and analysis of the image and video with the motion as the core, including: the modeling and characterization of the camera motion in the process of image generation, the motion estimation of the camera in the blurred image, and the image restoration. The multi-level representation of the motion in the video content record and the rapid analysis of the motion. The main work and innovation points of this paper can be summarized as follows: 1. In the process of image imaging, the deep analysis of the motion of the camera is carried out, and the model of the motion kernel is decomposed and optimized by itself. The decomposition model can reveal the intrinsic characteristics of the camera imaging, so as to observe the classical image de-blurring problem with a brand-new angle. In order to demonstrate the advantages of the characterization, we propose a trajectory random perturbation algorithm to optimize the motion kernel. In many examples, when we find that the existing deblurring algorithm falls into the local extreme value, our algorithm can obtain better de-blurring effect and correct blur kernel by independently optimizing the camera track. This feature of the camera motion can be more accurately described by using the high-bright area in the image-forming process, and it becomes feasible to combine the problem with the high-bright area for the very challenging scene of the night scene. In addition, we put forward a new frame that organically combines the motion kernel and the non-high bright region inferred from the high bright region to find a more accurate motion kernel. In addition, we propose a new functional motion kernel representation to accurately infer the motion kernel from the high bright region, We propose a new energy minimization equation that can automatically assign the extracted motion kernel to different areas for non-uniform deblurring. In the video content record, we focus on the analysis of the important motion information in the video content: the fall motion detection, in order to adapt to the real-time motion analysis in the video stream scene, that is, the real-time fall detection, we divide the motion information in the video content according to the "degree of difficulty", The action detection is carried out in a cascade manner, which is different from the traditional cascade framework, which can support the characteristics of different complexity. Through the cascading frame of this kind of mixing characteristic, our system can achieve a good compromise on the accuracy and efficiency. In addition, we designed the features we adopted, support the feature multiplexing and incremental updating, so that the video stream scene can be well supported. Finally, on the basis of the detection of the fall motion, we have developed so as to be able to support the general kinds of motion detection and to introduce more kinds of features so as to have a better improvement in the accuracy. This paper makes an in-depth analysis of the motion in the process of image generation and in the video content record. A large number of experiments show the effectiveness of the motion modeling of the camera and the high efficiency of rapid detection of motion layering in the video content.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41
[Abstract]:With the rapid development of multimedia technology and sensor technology, image and video, as the carrier of recording vision, have become more and more important to people's life, production and so on. In the last decade, the demand for recording visual information in various aspects has resulted in an explosive image and video. In these data, motion information plays an important role in both the generation and recording of images and videos, and can be said to cover various fields of image and video processing, and in the semanteme it also covers the motion understanding from the low-level motion track to the high-level motion. In particular, in the image generation process, the camera motion causes the blurring of the image, and the information entropy of the image is increased, and in order to be able to recover a clear sharp image, it is often necessary to accurately estimate the motion of the camera. during the recording of the image it is possible that the motion blur of the image content is caused by the rapid movement of the object or person. The motion blur generated in the recording process tends to be more information than the information provided by the clear image because it captures the dynamic object superposition. The video is generally considered to be a superposition of the images in time series, similarly, during the production of the video, the motion of the camera often results in the shaking and blurring of the picture, whether in the image quality and in the video. In the video content record, the motion information is often the reason of the existence of the video, and the analysis of its motion tends to pay more attention to the semantics of the higher layer. This paper studies the motion estimation and analysis of the image and video with the motion as the core, including: the modeling and characterization of the camera motion in the process of image generation, the motion estimation of the camera in the blurred image, and the image restoration. The multi-level representation of the motion in the video content record and the rapid analysis of the motion. The main work and innovation points of this paper can be summarized as follows: 1. In the process of image imaging, the deep analysis of the motion of the camera is carried out, and the model of the motion kernel is decomposed and optimized by itself. The decomposition model can reveal the intrinsic characteristics of the camera imaging, so as to observe the classical image de-blurring problem with a brand-new angle. In order to demonstrate the advantages of the characterization, we propose a trajectory random perturbation algorithm to optimize the motion kernel. In many examples, when we find that the existing deblurring algorithm falls into the local extreme value, our algorithm can obtain better de-blurring effect and correct blur kernel by independently optimizing the camera track. This feature of the camera motion can be more accurately described by using the high-bright area in the image-forming process, and it becomes feasible to combine the problem with the high-bright area for the very challenging scene of the night scene. In addition, we put forward a new frame that organically combines the motion kernel and the non-high bright region inferred from the high bright region to find a more accurate motion kernel. In addition, we propose a new functional motion kernel representation to accurately infer the motion kernel from the high bright region, We propose a new energy minimization equation that can automatically assign the extracted motion kernel to different areas for non-uniform deblurring. In the video content record, we focus on the analysis of the important motion information in the video content: the fall motion detection, in order to adapt to the real-time motion analysis in the video stream scene, that is, the real-time fall detection, we divide the motion information in the video content according to the "degree of difficulty", The action detection is carried out in a cascade manner, which is different from the traditional cascade framework, which can support the characteristics of different complexity. Through the cascading frame of this kind of mixing characteristic, our system can achieve a good compromise on the accuracy and efficiency. In addition, we designed the features we adopted, support the feature multiplexing and incremental updating, so that the video stream scene can be well supported. Finally, on the basis of the detection of the fall motion, we have developed so as to be able to support the general kinds of motion detection and to introduce more kinds of features so as to have a better improvement in the accuracy. This paper makes an in-depth analysis of the motion in the process of image generation and in the video content record. A large number of experiments show the effectiveness of the motion modeling of the camera and the high efficiency of rapid detection of motion layering in the video content.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41
【相似文献】
相关期刊论文 前10条
1 孙英皓;唐棣;;一种有效的交互式图像局部内容放大方法[J];计算机工程与应用;2012年12期
2 李瑞帅;余淑华;潘凯;王圆;;图像特征金字塔快速计算方法[J];电子世界;2014年07期
3 曾喜良;王金娟;;指纹分割的块图像梯度因子聚类法[J];计算机与数字工程;2008年07期
4 刘陈;王欣欣;李凤霞;赵相坤;;一种快速保边的图像对象分割方法[J];北京理工大学学报;2010年02期
5 朱薇;刘利刚;;图像适应算法中非冗余显著图的计算[J];中国图象图形学报;2011年08期
6 柳有权;吴宗胜;韩红雷;吴恩华;;线条增强的建筑物图像抽象画生成[J];计算机辅助设计与图形学学报;2013年09期
7 吴骏,唐红梅,肖志涛,贾志成;一种基于相位信息的图像对称性检测方法[J];信号处理;2004年01期
8 邵静;高隽;赵莹;张旭东;;一种基于图像固有维度的感知物体检测方法[J];仪器仪表学报;2008年04期
9 潘如如;高卫东;;高紧度机织物图像倾斜的自动纠正[J];纺织学报;2009年10期
10 刘贵喜,赵曙光,杨万海;基于梯度塔形分解的多传感器图像融合[J];光电子·激光;2001年03期
相关会议论文 前5条
1 张一鸣;刘亚t，
本文编号：2336262
本文链接：https://www.wllwen.com/shoufeilunwen/xxkjbs/2336262.html
