《动态背景下目标图像分割方法的研究421》由会员分享，可在线阅读，更多相关《动态背景下目标图像分割方法的研究421（117页珍藏版）》请在金锄头文库上搜索。
1、摘要摘 要随着图像处理理论和计算机技术的发展，动态图像的目标获取己经成为计算机科学、机器视觉、图像工程、模式识别、人工智能领域研究的重要课题。对于动态图像的分析是图像处理与计算机视觉领域一个十分活跃的分支和备受关注的前沿课题。动态目标图像获取的基本任务是从图像序列中识别出运动信息和背景信息，实现运动信息与复杂背景的精确的智能分割。它在军事、工业、智能交通、医学和多媒体等领域方面都具有重要的意义，发展应用前景广阔。在现代的智能交通及多媒体应用技术中，常常需要检测出运动的人体或车体，并将其与背景分离。即要求解决复杂背景下动态目标的分割问题。静态图像分割技术可以应用于序列图像的目标检测，目前的静态图
2、像分割技术大致可以提取出目标的轮廓，但边缘的细化问题仍然没有很好的解决。本文提出了针对该问题的解决方法，并把运动前景中运动目标和它投射的运动阴影分离开来，有效的抑制了阴影的影响，实现了动态目标的准确分割。动态图像序列的目标分割相对于静止图像来讲，所含的信息量要大很多，所以要选择合适的算法。考虑到动态图像序列有很多的复杂情况，分割算法要更加细化。动态图像序列存在着实时背景的变化，这就需要构造动态背景；在序列图像的各帧中，由于目标体的运动，会有个别帧目标边缘模糊，或出现与背景颜色相近的情况，因此在差分过程中容易出现目标体的局部缺失，针对此问题本文采用了分割逼近法对图像进行粗分和细分两次分割，建立了
3、人物图像边缘连通体再替换原图像的方法。首先，总体介绍了目标获取方法的各个步骤，包括目标的提取、背景的形成、阴影的分割。讨论了传统的图像分割算法和边缘检测算法，和边缘检测的评价标准。图像边缘检测是至今没有得到圆满解决的一类问题，通过追踪未知目标的轮廓而扫视这个物体。对一些传统的边缘检测算子，如：一阶Roberts算子、Sobel算子等等进行了原理分析，并将这些算子应用于目标图像，测试其边缘检测的效果。但这些传统边缘检测算子都过于简单，其去除噪声以及确定边缘位置等性能不是很理想。本文也对拉普拉斯算子进行了分析，经过分析测试其边缘检测效果好于一阶算子，但是仍然存在噪声。进一步讨论了图论的边缘检测方法
4、。对于图像分割，最基本的是阈值法，对灰度图象分割效果较好，而特征空间聚类法在彩色空间分割里应用较多。针对背景的扰动和光线的变化，为了准确提取运动目标，本文提出一种实用的方法就是对背景图像进行选择性动态更新，对背景图像中除掉运动目标对应位置之外的点进行自适应更新。同时讨论了阴影的形成和对目标分割的影响。 其次，对彩色空间内的目标图像的分割进行了专题研究，彩色图像分割中，彩色空间的选择是图像处理的难点。每种彩色空间都有着自己各自的优缺点，都在不同的领域发挥着重要的作用。一般的思想是具体问题具体处理。本文完成了RGB空间和HSV空间的目标分割，并在两个空间里分别研究了阴影的抑制和分割噪声的抑制。然后
5、，将图像分割的目光转移到了基于动态目标的研究上来。因为光线、天气等条件的变化引起背景的改变，在这种情况下，只有依据一定的模型来更新背景图像，才能准确提取运动目标，本文利用帧间局部背景构造全背景的方法，实现背景的实时更新。首先在图像序列中找出一帧静态背景，然后利用帧间的更新背景不断替换前面的背景。这样在分割使用时能保证使用的背景是离被差分帧最近，从而达到误差最小的目的。并采用分割逼近算法勾勒出目标的轮廓，实现对目标的分割。另外，由于被阴影覆盖的背景与没有阴影的背景区域的颜色差别也很大，很难将真正的运动目标和运动阴影区分开，因此在分割出的运动目标中包括运动阴影区域，阴影去除对于准确的提取出运动目标
6、至关重要。本文对现有的阴影分割方法进行了分析与总结，指出了这些方法用于阴影分割时存在的问题，针对不同的应用要求，研究了基于HSV颜色空间的阴影去除方法和基于一阶梯度的阴影去除方法。最后，采用非线性视频编辑的技术，用Promiere Pro软件以一定的帧速率将每帧目标图片连接成视频文件，并与背景进行了合成。关键词 阴影去除，边缘检测，图像分割，目标恢复，动态背景- I -AbstractAbstractWith the development of image processing theory and computer technology, how to acquire the object
7、 of moving image has already been one important topic in the field of computer science, machine vision, image engineering, pattern recognition and artificial intelligence. The analysis to the dynamic image is an active branch and also a leading topic being paid great attention in the field of image
8、processing and computer vision. The basic task of dynamic object is to recognize the motion information and background information from the image sequence, then, segment the motion information and background information intellectualized and accurately. Such segment is of important meaning and wide d
9、evelopment in the future. In todays intelligent transport system and multimedia application technique, the moving body or car must be detected and separated from the background. That is the segmention problem of dynamic objects from static background. Static image can be applied to detecting the obj
10、ect of image sequence. The outline of the object can be approximately extracted with current static image segment technology, but the problem of edge thinning has not been solved in the computer vision field. To solve such problem, several resolving is brought forward in this dissertation. Frothier,
11、 in this dissertation, the shadow is separated from the object and restrain the influence of the shadow, which segment the motion image exactly.Comparing with the static image, the motion image is of its own feature, which is that it is of abundant motion information. As with such feature, the motio
12、n image sequence is of its own low cost path algorithm. As there are so many complex cases, the segmentaion algorithm should be thinner. The background is changing in the sequences of dynamic images, so the dynamic background should be constructed. As the result of the moving object, there may be fu
13、zzy edge in some frames or similar color with background, so the object will be partial dedicit. To solve such problem, the method is brought forward, which is based on segment aproaching method and edge extracion to replace original images.Firstly, each step of acquiring object is generally introdu
14、ced, including extracting object, forming background, segmenting shadow. The traditional image segment algorithm, edge detection algorithm and the assessment criterion of edge detection are also discussed in this dissertation. The image edge detection has not been completely resolved. In this disser
15、tation, the principle of traditional edge detection operator, such as Roberts operator, Sobel operator, then, test the effect of detecting edge by applying these operators to the object image. The traditional edge detection operator is so simple that the effect of removing noise and determine edge l
16、ocation is undead. Also, Laplace operator is analyzed in this dissertation. Although, the effect of detecting edge with Laplace operator is better than first order operator, there is still noise. Moreover, the method of detecting edge based on graph theory is discussed. To the image segment, the basic method is based on threshold which is fit to gray image segment. Correspondingly, the cluster method based on feature space is usually used in color
