项目名称: 基于在线判别学习的鲁棒视觉跟踪算法研究
项目编号: No.61203268
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 自动化学科
项目作者: 查宇飞
作者单位: 中国人民解放军空军工程大学
项目金额: 27万元
中文摘要: 在复杂背景下持续跟踪不断变化的目标,对跟踪算法的鲁棒性提出了严峻挑战。本项目拟将视觉跟踪看作在线判别学习问题来处理,在复杂背景中利用分类器区分变化的目标,在线更新分类器减少误差积累,实现对感兴趣目标的鲁棒跟踪。针对在线判别学习跟踪系统中的目标表示、在线学习和样本选择三个关键问题展开研究,通过融入目标结构信息的局部不变特征,构造目标表示的有效特征集,提取其本质信息;在线学习不仅利用标记样本来最大化目标和背景之间的分类间隔,而且还利用未标记样本来逼近目标和背景的内部结构,能够在复杂背景中区分出不断变化的目标;同时利用目标的时空连续性和候选样本的置信度,约束样本的选择,避免错误样本所引起的跟踪失败;最终实现对感兴趣目标的鲁棒跟踪。本项目的开展和预期成果将为视觉跟踪算法提供新思路和新方法,是一项既有理论意义又有广阔应用前景的应用基础研究课题。
中文关键词: 目标跟踪;相关滤波;稀疏编码;密度SIFT;局部图像块
英文摘要: There is a severe challenge for robust tracking algorithms to track the changed object persistently in the complex scene. The project intends to consider the visual tracking problem as an on-line discriminative learning problem. In order to track the object robustly, the proposed algorithm will identify the changed object from the complicated background by the classifier, which is updated on-inline to decrease the errors accumulation. We will focus on three problems of online tracking: object representation, online learning and samples selection. The effective feature set is constructed by local invariant structured features to extract the object intrinsic information. On-line learning not only unities the labeled samples to maximize the margin of the object and background, but also obtains the internal structure by the unlabelled samples. In order to avoid failed tracking by the error samples, the sample space-time coherence and the confidence are used to constrain the training set selection. The implementation and the expected results of the project will provide new ideas and new methods for the visual tracking algorithm, which has a broad theoretical basis for the application of the prospect research.
英文关键词: object tracking;correlation filter;sparse coding;dense SIFT;local patch