权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Spotting recognition and tracking of an object in a time-varying image using 2-dimensional Continuous Dynamic Programming

使用二维连续动态规划对时变图像中的对象进行定位识别和跟踪

基本信息

批准号：
16500115
负责人：
OKA Ryuichi
金额：
$ 2.11万
依托单位：
The University of Aizu
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (C)
财政年份：
2004
资助国家：
日本
起止时间：
2004 至 2006
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/en/grant/KAKENHI-PROJECT-16500115/
关键词：
recognition of time-varying image tracking spotting matching range image extraction non-linear registration retrieval of sequence pattern understanding of motion セグメンテーション移動物体認識連続DP ジェスチャ認識

项目摘要

Recognition of a deformable object in a time-varying image is one of difficult problems to be solved in image processing field. The difficulty is divided into four parts as follows : (1) segmentation of the target object in a time-varying image, (2) recognition of the deformable object in each frame image, (3) pixel-wise tracking of object image in a stream of image sequence, (4) recognition of the motion of object.. The proposed method solved each of the four problems described above. The first one is solved by successively application of so-called 2-dimensional Continuous Dynamic Programming (2DCDP) assuming a segmented image area of target object at starting time. The second is also simultaneously carried out during the application of 2DCDP. The pixel correspondence between a reference image and the spotted image is obtained by application of 2DCDP. Concerning the third, therefore, successively updating the reference image by using the spotted image during application of 2DCDP for each frame image of a time-varying image realizes the pixel-wise tracking of object image. The feature of motion of the target object can be extracted by using tracing of pixel-wise tracking because each pixel is belong to the part of object determined in advance as significant one. The improvement of accuracy of correspondence between pixels is realized by obtaining more precise spatial monotone and contiguous characteristics. Moreover we used pixel correspondence between successive images to make a 3-dimensional image of the target object by a single and moving camera. This method is different from the conventional methods like optical flow or factorization because our method is independent from the moving background. We also extended some algorithms based on Continuous Dynamic Programming for a wide application such as song wave retrieval which confirmed the applicability of 2DCDP.

时变图像中形变目标的识别是图像处理领域亟待解决的难题之一。其难点分为四个部分：（1）时变图像中目标物体的分割，（2）每帧图像中变形物体的识别，（3）图像序列中物体图像的逐像素跟踪，（4）物体运动的识别。所提出的方法解决了上述四个问题中的每一个。第一个是解决了连续应用所谓的2维连续动态规划（2DCDP）假设在开始时间的目标对象的分割图像区域。第二个也是在应用2DCDP的同时进行的。通过应用2DCDP获得参考图像和斑点图像之间的像素对应。关于第三，因此，在对时变图像的每帧图像应用2DCDP期间通过使用斑点图像连续更新参考图像实现了对象图像的逐像素跟踪。由于每个像素都属于预先确定为重要像素的目标部分，因此可以使用逐像素跟踪的跟踪来提取目标的运动特征。通过获得更精确的空间单调性和连续性特征，实现了像素间对应精度的提高。此外，我们使用连续图像之间的像素对应关系，使一个单一的和移动的摄像机的目标对象的三维图像。该方法与传统的光流法、因子分解法等方法不同，因为该方法与运动背景无关。我们还将一些基于连续动态规划的算法进行了扩展，使其应用于歌曲波形检索等领域，证实了2DCDP的适用性。