复杂背景下的目标实时分割与检测
文献类型:学位论文
作者 | 吴晓雨 |
学位类别 | 工学博士 |
答辩日期 | 2009-06-02 |
授予单位 | 中国科学院研究生院 |
授予地点 | 中国科学院自动化研究所 |
导师 | 王阳生 |
关键词 | 交互式分割 实时目标分割 实时目标检测 背景建模 图切 复杂背景 interactive segmentation real-time object segmentation real-time object detection background modeling graph cut complex scene |
其他题名 | Real-time object segmentation and detection in complex scenes |
学位专业 | 计算机应用技术 |
中文摘要 | 复杂背景下的目标实时分割与检测技术是计算机视觉领域的一个重要研究方向,在人机交互、智能监控和虚拟现实等领域具有广泛的应用前景。本文以视频中的目标实时分割与检测技术为研究主体,对其中的一些关键问题进行了探索和研究。 论文的主要工作归纳如下: (1) 提出了基于区域与像素级的交互式图切分割算法。该方法将图像分割成几个区域,借助用户交互信息,建立了基于区域级的赋权图,利用区域合并的分水岭结果及邻域像素的颜色差构造Gibbs能量函数,通过图切算法(graph cut)实现前景/背景的区域分割;同时为了得到准确的前景/背景边界,在前景的边界带进行了基于像素级的二次图切分割。实验结果表明了这种交互式分割方法的有效性。 (2) 提出了特征背景模型与高斯模型自适应融合的背景模型。该模型基于主成份分析方法建立了像素亮度信息的特征背景及色度信息的高斯模型,并设定了基于光照变化的融合机制。同时对模型分割的结果进行了阴影去除及后处理。实验结果表明融合的背景模型综合了特征背景模型和高斯模型的优势,提高了目标的实时分割精度。 (3) 提出了基于时空连续性的实时目标分割方法。该方法首先利用提出的融合背景模型将当前帧图像分割成前景、背景和未知标签三类像素集,然后在动态图切框架下,根据颜色和对比度信息构造目标能量函数,设计了基于时间连续性信息的融合背景模型和全局前景模型的数据项,提出了基于局部二值模式(lbp)的对比度平滑项。最后通过图切算法极小化能量函数求取所有像素的二值标签。采用边界平滑和alpha值估计等后处理方法,使分割出的前景目标无缝地融合到虚拟背景中。实验结果表明该方法能较好地将复杂背景中的目标实时分割出来并真实地合成到虚拟背景中。 (4) 提出了基于梯度方向直方图特征的手势检测方法。该方法利用梯度方向直方图(hog)特征提取不同手势的形状信息,用级联的Adaboost学习算法构造检测器。在检测过程中,根据前一帧检测结果确立感兴趣区域,从而实现实时的多角度手势检测。 |
英文摘要 | As one of the important research area, real-time object segmentation and detection can be used in many fields like human-computer interaction, intelligent surveillance, virtual-reality, etc. This paper focuses on real-time object segmentation and detection. The main contributions of the paper are as follows: (1) A two-stage method combing the region-level and pixel-level processing is proposed to segment the foreground object with the help of user's interaction. At first, watershed algorithm based on color feature is chosen to cluster the image into different regions. With the hint of user's interactive, a weighted graph is developed to minimize the Gibbs energy function to assign the binary labels to all regions by graph cut algorithm. Finally, a pixel-level graph around the foreground boundary band is developed to obtain the accurate foreground edge by pixel based graph cut algorithm. Experimental results demonstrate our method's efficiency and effectiveness. (2) A fused background model that combines the eigenbackground with Gaussian models is proposed. We adopt the eigenspace model to build the intensity information for each pixel. Uni-modal Gaussian density methods with less computational cost are used to describe color information for each pixel. An adaptive strategy is used to integrate the two models. Using the fused background model, we subtract the background from the current video frame to obtain the foreground object. Shadow removal based on chroma color method and post-processing are discussed in the end. Experimental results prove that our model is robust to noise and illumination change due to inheriting eigenbackground and Gaussian's model advantages to improve the segmentation results. (3) An automatic foreground segmentation system based on spatial-temporal continuity and dynamic graph cut algorithm is developed. Firstly, the current frame is separated into the definitely foreground, definitely background and unknown pixels by our background model mentioned above. Second, the fusion background model based on temporal continuity and global foreground model are used to set the data term in the Gibbs energy function. The smooth term is calculated by local binary pattern contrast. In order to accelerate the segmentation computation process, dynamic graph cut algorithm is implemented to minimize the energy function and obtain binary labelings for all the pixels. In addition, we design a frequency filter to smooth the foreground boundary a... |
语种 | 中文 |
其他标识符 | 200618014629085 |
源URL | [http://ir.ia.ac.cn/handle/173211/6211] ![]() |
专题 | 毕业生_博士学位论文 |
推荐引用方式 GB/T 7714 | 吴晓雨. 复杂背景下的目标实时分割与检测[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2009. |
入库方式: OAI收割
来源:自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。