项目名称: 多视点视频追踪问题的研究
项目编号: No.11301239
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 数理科学和化学
项目作者: 汪帆
作者单位: 兰州大学
项目金额: 22万元
中文摘要: 在本项目中,我们提出一种先混合多个视角的视频数据,再检测背景和前景信息从而进行目标追踪的多视角目标追踪算法。现在流行的使用数个摄像头视角进行追踪的算法都是先检测不同镜头内多个目标的特征信息进行前景背景分析,然后将得到的信息混合起来进行目标追踪。然而这种在单个摄像头视角内检测信息的方法受环境影响很大,例如发光物体的反射和阴影,以及各个目标之间有无遮挡的情况等。在我们的项目中,首先通过将各个视点的信息投影到参考平面以及平行于参考平面的多个平行平面的投影方法混合各个视角内信息形成一个四阶张量,然后使用增长的张量子空间算法建立背景模型进行前景分割,最后对目标区域应用贝叶斯定律以及粒子滤波等方法来进行追踪,我们将对比现在流行的各种视频追踪算法的结果来验证我们的结论。项目预期我们的方法可以在处理阴影,遮挡,拥挤环境中多个目标的追踪等困难问题上能取得一定的研究成果,为进一步的研究打下一个良好的基础。
中文关键词: 多视点;目标追踪;前景分割;乘法噪音;凸优化
英文摘要: In this project,we present a novel fuse-before-detect algorithm for multi-view object tracking via fourth order tensor learning.By using several camera views, most of the existing algorithms first detect the various object features for each view and then fuse the data together for tracking.However,this kind of single view foreground segmentation algorithm always suffers from various environmental problems,such as reflection and shadow induced by shiny objects,especially floor and wall.These segmentation errors reduce the accuracy of the multi-view tracking algorithms.In our project,we first fuse multi-view camera data to a fourth-order tensor through multiple parallelized planes projection.An incremental tensor learning algorithm is then employed to perform foreground segmentation in the fused tensor data.By collecting all the information from different views,this approach could restrain the specific environmental effects in each view and give better segmentation results.Then we use the incremental tensor subspace learning algorithm to appearance-based object tracking.A Markov model with hidden state variables is used for motion estimation.An affine image warping is applied to model the object motion between two consecutive frames.Given a set of observed image regions the posterior probability of the object stat
英文关键词: multi-view;object tracking;foreground segmentation;multiplicative noise;convex optimization