基于超像素的判别式目标主动轮廓跟踪

项目名称： 基于超像素的判别式目标主动轮廓跟踪

项目编号： No.61472063

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 周雪

作者单位： 电子科技大学

项目金额： 80万元

中文摘要： 现有的主动轮廓跟踪方法研究主要基于底层视觉特征-像素，但其易受噪音干扰，由此产生的跟踪准确性和鲁棒性的问题已经成为制约该类方法的瓶颈。本课题针对主动轮廓跟踪开展系统研究，在归纳现有算法局限性的基础上，拟将一种有效的中层视觉特征-超像素引入到判别式主动轮廓跟踪框架中，直接将轮廓进化方程中的速度函数建模成由表观和形状先验决定的判别因子项。本课题旨在引入机器学习等理论，结合模式识别技术，从选择图像基本描述单元和挖掘高层先验知识两方面考虑提高算法性能。研究内容包括：(1)快速有效的超像素分割；(2)基于超像素的多模态判别式表观建模；(3)融合多尺度超像素的表观特征学习；(4)基于稀疏表示的形状建模。本课题的研究目标是借鉴并利用多学科的理论和方法，为提高跟踪准确性和鲁棒性，构建基于超像素的判别式主动轮廓跟踪基本理论与方法，并探索其在实际中的应用，最终提供更为有效、更为便捷的主动轮廓跟踪解决方案。

中文关键词： 计算机视觉；主动轮廓模型；超像素；判别式跟踪；形状先验

英文摘要： Current active contour-based tracking methods mainly build the pixel-wise models which are sensitive to noise disturbance. Thus this kind of methods are subject to low effectiveness and robustness.In this proposal, with the summarization of previous approaches, we believe it is a good solution by introducing mid level visual cue-superpixel into the discriminative active contour tracking framework. We directly formulate the speed function in the contour evolution equation as a discriminative factor determined by the appearance and shape priors.By introducing machine learning theory and combining pattern recognition technique, this proposal aims to focus on selecting the basic effective image representation unit and mining high level prior knowledge in order to improve the performance of algorithms. The research content includes: (1) Fast and effective superpixel segmentation; (2) Building superpixel-based multimodal discriminative appearance model; (3) Appearance representation learning with fusion of multiscale superpixels; (4) Modeling Sparse representation-based shape priors. Our final objective is to utilize theories from multiple research areas and improve the effectiveness and practicability of active contour-based tracking with a better solution.

英文关键词： Computer vision;Active contour model;Superpixel;Discriminative tracking;Shape priors