题名 | 动态纹理建模及其在视频检测中的应用 |
作者 | 王谦 |
学位类别 | 工学博士 |
答辩日期 | 2007-06-08 |
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 | 谭铁牛 |
关键词 | 动态纹理 视频检测 高斯过程 视频跟踪 Dynamic Texture Video detection Gaussian Process Video Tracking |
其他题名 | Dynamic texture modeling and its applications in video detection |
学位专业 | 模式识别与智能系统 |
中文摘要 | 本文的工作以建立视频检测平台为目标,并以此为中心对肤色检测、动态纹理建模、视频跟踪、镜头切换检测等问题进行了初步探讨,重点研究核心的动态纹理建模方法并把这种方法应用于敏感视频检测中,取得了初步的效果。论文的主要工作和贡献如下: 1.建立了初步的视频检测实验平台,这个平台用于检测网络上的敏感视频,包括肤色检测、关键帧抽取、敏感度判据等算法。 2.将线性动态纹理模型应用于镜头切换检测和降低视频检测的误判。本文求解出线性动态系统的转移噪声协方差阵,并将其应用于敏感视频检测中,实验结果表明这种方法可以明显降低敏感视频的误判比率。本文还将转移噪声协方差阵和图像频率信息应用于镜头切换检测,在实验中获得了初步的识别结果。在线性动态纹理模型的基础上,本文提出了变分PCA动态纹理模型。 3.从回归模型的角度出发,本文提出了一种基于高斯过程线性动态模型的动态纹理建模方法,然后用高斯过程预测的方法来合成动态纹理,并将其与传统的方法进行了比较。 4.本文采用动态纹理的思想来做视频跟踪,并利用视频跟踪结果来降低敏感视频检测的误判。在获得视频中的跟踪轮廓后,本文将视频检测精确到运动轮廓区域上。实验结果表明这种方法可有效降低误判率。 |
英文摘要 | Aiming to build a platform for video detection, we investigate some of the key issues related to skin detection, dynamic texture modeling, video tracking and shot changing detection. Meanwhile, we apply dynamic texture modeling for objectionable video detection and obtain some improvements in decreasing the false alarm. The main contributions of our work reported in this thesis are as follows: 1.An experimental platform of video detection is built where one can filter the video with objectionable content on Internet. Some algorithms have been realized, including skin detection, key frame extraction and objectionable degree estimation. 2.Linear dynamic texture model is used to detect shot changes and decrease the false alarm of objectionable video detection. Using EM algorithm, the transfer covariance matrix is computed. Our experiments show that the false alarm obviously decreases in objectionable video detection. We also use this method in shot changing detection and have preliminary results. We propose Variational PCA (VPCA) dynamic texture model based on linear model. Compared with classical linear models, VPCA has a smaller reconstruction error. 3.Gaussian process linear dynamic model (GPLDM) is proposed for dynamic texture modeling and synthesis, using a three-level framework. Experimental results indicate that the latent structure of GPLDM is much smoother than GPLVM and GPDM. 4.We treat the task of motion tracking as a special case of dynamic texture prediction. Gaussian process linear dynamic model is used to track human beings. After motion tracking, we compute skin region area based on motion contours, which is different from previous method, which compute skin region area based on whole image. The skin ratio is motion skin region area divided by image region area. Our experiments show that the false alarm is lower than that of original method. |
语种 | 中文 |
其他标识符 | 200118014604888 |
内容类型 | 学位论文 |
源URL | [http://ir.ia.ac.cn/handle/173211/6012] ![]() |
专题 | 毕业生_博士学位论文 |
推荐引用方式 GB/T 7714 | 王谦. 动态纹理建模及其在视频检测中的应用[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2007. |
个性服务 |
查看访问统计 |
相关权益政策 |
暂无数据 |
收藏/分享 |
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。
修改评论