分层视觉模型及表观复杂变化的视觉目标跟踪方法研究

项目名称： 分层视觉模型及表观复杂变化的视觉目标跟踪方法研究

项目编号： No.61300099

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 王玉茹

作者单位： 东北师范大学

项目金额： 22万元

中文摘要： 基于视频图像序列的目标跟踪是计算机视觉的热点问题，然而目前大多数的应用条件局限于目标外观的简单变化，这使得许多实际应用受到限制。本申请针对复杂场景下，目标存在复杂结构和外观变化这一制约跟踪技术付诸更广泛应用的难题，研究鲁棒的分层视觉模型和高效精确的基于分区域采样的跟踪算法。首先，将目标整体表示为多个子块的集合，构建联合局部视觉特征、邻域结构特征和全局视觉特征的自适应分层视觉模型，以对目标复杂的表观进行鲁棒建模；其次，对目标的状态空间建立分区域采样概率模型，优化粒子滤波器的粒子采样过程，从而处理具有复杂结构变化的目标局部子块突变和平滑运动共存的问题，并达到高效精确的跟踪；最后，在视频序列上获取目标运动完整的时空轨迹。本研究成果适用于具有各种复杂场景和目标外观的视频跟踪问题，将为更高层次的视频分析和理解提供有效的信息和决策支持，对于推动视频运动分析的发展和实用化具有重要的理论意义和实用价值。

中文关键词： 计算机视觉；视频目标跟踪；分层视觉模型；贝叶斯概率框架；集成跟踪

英文摘要： Visual tracking is a hot issue in computer vision, but in current, most of the applications are limited to the condition of target's appearance's simple changes. In real applications, both the complex environment and target's rapid structural and appearance changes present the current tracking techniques with serious difficulties to be applied in wider range of applications. Focusing on the above problem, this project will make major research on developing a robust layered visual model and a regional sampling based efficient and accurate tracking algorithm. Firstly, the target is described as a set of patches, and a layered visual model is constructed by integrating local visual, neighbor structural, and global visual features, to provide a robust representation of target's complex appearance. Secondly, a regional sampling probability model is designed in target's state space to optimize particles' sampling procedure, so as to deal with local patches' smooth and abrupt motions and realize an effective and accurate tracking. Finally, target's whole spatial-temporal trajectory is output in the video sequence. The results of this research is applicable for the tracking problems with various scenes and appearances, will provide information and decision support for visual analysis and understanding in a higher level,

英文关键词： Computer Vision；Visual Object Tracking；Layered Visual Model；Bayesian Probability Framework；Ensemble Tracking