判别式表观建模方法

项目名称： 判别式表观建模方法

项目编号： No.61472036

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： 马波

作者单位： 北京理工大学

项目金额： 80万元

中文摘要： 表观建模是视觉跟踪研究的核心，稀疏跟踪和在线学习是当前研究热点。这一领域还存在如下问题从而制约了跟踪的准确性和鲁棒性：稀疏特征的池化方法只考虑了码本系数的低阶特征，不能充分建模目标表观；大多数稀疏跟踪算法基于样本重建误差建模似然函数，不能充分提取最有利于跟踪的判别信息；表观模型在线更新时样本选取缺乏原理性准则等等。针对这些问题，基于主动学习、非线性学习和稀疏编码理论，本项目拟研究：基于锚点学习的半监督稀疏跟踪；结构化局部稀疏视觉描述子和基于Fisher向量的局部稀疏视觉描述子；采用主动学习理论的公理化样本选取等内容。项目实施会为智能监控等视觉目标跟踪应用提供关键技术与核心算法。

中文关键词： 视觉跟踪；判别式表观建模；主动学习；稀疏编码；非线性学习

英文摘要： Appearance modeling is the key component in visual tracking,and sparse tracking and online learning are two hot topics. Several challenges in this field haven't been addressed that affect tracking accuracy and robustness. When pooling sparse codes,most methods only consider their low order information, and fail to model target appearance effectively.Most sparse tracking methods model likelihood function using sample reconstuction error, and fail to extract most disriminative information to perform tracking. Current methods adopt heuristic strategies rather than principled ones to online update apearance models. We start from active learning theory, sparse coding theory and nonliear learning theory,and make the following proposals that are intended to address the aforementioned issues: semi-supervised sparse tracking methods using achor points learning, structured local sparse visual representation and fisher vector based local sparse visual representation, and a principled sample selection method for online appearance model updating using active learning. If this proposal is implemented, we expect it would provide key techniques and algorithms for viusal tracking applications like intelligent surveillance.

英文关键词： visual tracking;discriminative appearance modeling;active learning;sparse coding;nonlinear learning