支持增量式稀疏编码的在线协同目标跟踪研究

项目名称： 支持增量式稀疏编码的在线协同目标跟踪研究

项目编号： No.61302156

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 无线电电子学、电信技术

项目作者： 韩光

作者单位： 南京邮电大学

项目金额： 24万元

中文摘要： 基于视频的目标跟踪由于其在民用和军用领域具有极其重要的应用价值，因而目标跟踪现在仍然是计算机视觉中的研究热点之一。由于目标图像易受到姿态、形状、光照的变化以及遮挡等因素的影响，想要实现稳定可靠的目标跟踪仍然有许多亟待解决的问题，如目标易丢失，样本数据的不平衡性等。针对以上问题，本项目将以基于增量式稀疏编码的跟踪，基于主动选择的SVM检测和基于协同机制的在线学习为攻关内容对复杂状况下的目标跟踪问题进行基础研究探讨。本项目的创新之处：拟提出基于结合1、2范数的稀疏编码及其特征选择算法和在线的目标子空间更新方法；考虑设计利用K均值聚类选取最有代表性负样本的策略来提高SVM检测器的学习速度和识别能力；拟利用协同机制解决跟踪器与检测器所得样本不一致的问题；提出的目标跟踪框架能实现对跟踪器与检测器的同步更新。本课题将有利推动模式识别、机器学习和视频监控理论、技术和应用。

中文关键词： 目标跟踪；稀疏表示；相关滤波；在线学习；协同跟踪

英文摘要： Video object tracking now is still one of the research hotspots in computer vision，because it has very important application value in military and civilian fields. The object image is always affected by the changes of pose, shape and illumination, occlusions and other factors，there is still many problems to be solved for achieving stable and ralisble object tracking, such as the object being lost, the imbalance of sample data. Aiming at overcoming the above challenge, this project mainly discusses the tracking based on the incremental saprse coding, detection based on active SVM as well as online learning based on cooperative mechanism. The innovation of the project is: (1) to propose a robustly sparse coding algorithm, which includes the constrcutions of a sparse model via joint 1、2 norms and a feature selection approach, object subspace updating method based on online learning; (2) one strategie is used to improve the learning speed and recognition ability of SVM algorithm, it is to use a dynamic clustering for selecting the best representative negative samples; (3) to solve the inconsistent problem on the samples obtained from the tracker and detector, a cooperative mechanism is used; (4) the proposed framework of object tracking can realize synchronous updating on the tracker and detector. This study will p

英文关键词： Object tracking；Sparse representation；Correlation filter；Online learning；Cooperative tracking