复杂场景下目标跟踪的表观建模研究

项目名称： 复杂场景下目标跟踪的表观建模研究

项目编号： No.61503388

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 其他

项目作者： 丁建伟

作者单位： 中国人民公安大学

项目金额： 20万元

中文摘要： 表观建模是鲁棒目标跟踪的核心问题。针对现有表观建模方法在复杂场景下存在的不足，本项目拟从视觉描述和模型学习两方面进行系统的研究。通过研究高效的基于局部的拓扑结构模型和融合底层特征与高层属性信息，建立更有效的视觉描述方式。基于半监督模型学习框架，探索将先验知识与在线训练的分类器有效融合，减少漂移发生的概率，提升跟踪的准确性。通过构造跟踪结果与样本提取评估模型，改善错误样本更新导致的模型恶化问题，提升算法在完全遮挡和跟踪失败等极端情况下的鲁棒性。

中文关键词： 目标跟踪；复杂场景；表观建模

英文摘要： Appearance modeling is the critical problem of robust object tracking. In order to overcome the shortcomings of current appearance modeling methods in complex scenes, we are intended to do systematically research from visual representation and model learning in this proposal. More effective visual representation ways are proposed by studying efficient local topological structure model and fusing low level features and high level attributes information. We will study how to effectively fuse prior knowledge and online learned classifier in semi-supervised model learning architecture, which can reduce occurrence probability of drift and promote tracking accuracy. Evaluation models of tracking results and samples extraction are generated to improve model deterioration for wrong samples updating, and promote robustness of algorithm in extreme cases, such as fully occlusion and tracking failures.

英文关键词： object tracking;complex scenes;appearance modeling