欢迎您！东篱公司
退出
-
申报数据库
-
立项数据库
-
成果数据库
-
项目获奖数据库
中文摘要：
通过构建基于超像素的图作为视觉表示引入超像素间的空间信息.采用基于图模型的流形排序作为显著性检测方法得到第一阶段每个超像素的显著性,判别式表观模型则通过基于中层特征的分类器进行判别并利用空间信息对分类结果进行调整,将流形排序和分类结果结合作为先验信息选择随机游走种子点.结合随机游走得到的第二阶段的显著值和分类结果,最终得到当前帧的置信图.在置信图的基础上,采用积分图方法快速计算得到候选的观测值,将观测值最大的候选作为跟踪结果.在数据集上的实验结果表明,该方法可以有效处理快速运动和形变等问题,从而实现复杂背景下鲁棒的目标跟踪.
英文摘要：
In this work,we focus on short-term single object tracking,which is the most general type of tracking problems.Numerous significant trackers have been proposed over the past few decades.As we can see from these trackers,the methods that adopt the mid-level representation have shown their superiority over other approaches in dealing with challenging factors like partial occlusion.However,most of their representation model lack the spatial information,which usually leads to poor robustness in object tracking.As a popular middle level representation,superpixels are semantically meaningful and much more homogeneous than randomly selected square patches.In this work,we construct a graph based on superpixels to introduce spatial information.A graph-based saliency detection model,which uses manifold ranking to compute the saliency scores for each superpixel for the first stage,is combined with a discriminative model,which trains a classifier to classify the candidate superpixels as target or nontarget.The saliency scores and the classification result adjusted by the spatial information are then used to select the seeds for random walk as prior knowledge,which makes the result of saliency detection more relevant to the target object.Combining the classification result and the saliency scores computed by the second stage random walk,a confidence map is achieved,based on which candidates are fast ranked utilizing integral graph method.The top-rankedcandidate is regarded as the target object.In order to evaluate our approach,we compare our results with other trackers along with some analysis.The experimental results on visual tracking benchmark dataset demonstrate that our approach is effective for fast motion,partial occlusion and background clutter in tracking,thus realizes desirable and robust tracking performance under complex conditions.
同期刊论文项目
同项目期刊论文
期刊信息
位置：