项目名称: 多目标跟踪中的注意模型研究
项目编号: No.61473031
项目类型: 面上项目
立项/批准年度: 2015
项目学科: 其他
项目作者: 邹琪
作者单位: 北京交通大学
项目金额: 80万元
中文摘要: 人类视觉如何在被跟踪的多个目标中分配注意,一直是认知和神经科学非常关注的问题。它既具有自底向上和自顶向下注意的共性,还体现了多目标跟踪的特性。借鉴人类视觉在多目标跟踪中注意机理,利用视频跟踪中的眼动的研究成果,研究复杂动态环境的注意模型。一方面,我们从动态变化的场景中学习到空间布局等场景上下文;也将学习目标有区分度的表观特征和高阶运动模型,用于形成对场景和特定任务的有效表达。将学习到的先验作为知识和熟悉性,推广到未曾学习过的、新出现的目标。另一方面,借鉴多目标跟踪的注意理论,通过时空组织形成目标组,利用在线随机森林学习目标与任务的相关性,构建多目标跟踪的自顶向下注意模型。将所构建的注意模型应用在基于高阶运动量和多维数据关联的多目标跟踪中,解决遮挡和场景动态变化的干扰,实现鲁棒的跟踪。预计这一成果将缩小跟踪计算模型与人类视觉跟踪之间的鸿沟,加深我们对自顶向下注意的理解
中文关键词: 多目标跟踪;自顶向下注意;眼动;动态感知;时空组织
英文摘要: How visual attention is deployed in multi-object tracking(MOT)? It has been noticed in perceptual and neural science for a long time. It has common properties of bottom-up and top-down attention, and has properties specific to MOT.Inspired by attention mechanism of MOT in human vision, we will biuld computational attention models in complex dynamic scenes.On the one hand, we learn scene context in dynamic scenes, and discriminative appearance and high-order motion models of objects, to represent scenes and tasks efficiently. On the other hand,based on attention and eye movement theories and learning relations of objects with tasks by online random forest, we will build a top-down attention model. Applying the attention model in MOT which combines spatial-temporal grouping and multidimentional data association,coping with occlusion and distractors from dynamic background, we try to realize robust tracking. The research is expected to reduce gaps between computational tracking models and human visual tracking,and at the same time deepen our understanding to top-down attention.
英文关键词: multiple object tracking;top-down attention;eye movememts;dynamic perception;spatial temporal grouping