中国科学院机构知识库网格
Chinese Academy of Sciences Institutional Repositories Grid
复杂场景下视觉单目标跟踪算法研究

文献类型:学位论文

作者白延成
学位类别工学博士
答辩日期2014-05-22
授予单位中国科学院大学
授予地点中国科学院自动化研究所
导师唐明
关键词目标跟踪 表观模型 排序学习 弱监督学习 特征融合 多任务学习 Object Tracking Appearance Model Learning to Rank Weakly Supervised Learning Feature Combination Multi-task Learning
其他题名Research on Visual Single Object Tracking
学位专业模式识别与智能系统
中文摘要视觉目标跟踪是计算机视觉领域中的研究热点之一。目标跟踪在智能监控、人机交互、虚拟现实等领域具有广阔的应用前景。经过几十年的研究,目标跟踪技术有了长足的发展,但是复杂场景下对任意目标进行稳定、准确地跟踪仍然存在诸多理论和技术问题亟需解决,特别是复杂场景存在的光照变化、复杂背景、平面外选择、严重遮挡等诸多难点。 一个典型的目标跟踪系统由三部分组成:1)表观模型主要用于评估特定区域与目标似然程度;2)运动模型关联目标在不同帧中出现的位置;3)搜索策略是指如何有效地确定目标在当前帧中最有可能出现的位置。其中表观模型是跟踪系统的核心,是目标跟踪关键之所在。因此,本文的研究内容主要是设计鲁棒的表观模型以提高跟踪系统的鲁棒性。论文的主要工作和贡献如下: 1. 为了减少误标记样本对表观模型更新的影响,提出了基于排序学习的目标表观模型跟踪算法。该算法利用样本之间的相对优劣很容易判断的特点,将跟踪问题转化为排序学习问题,并利用Ranking SVM学习样本之间的相对序关系。实验表明该算法能较好地处理目标的姿势变化、光照变化等问题。 2. 提出了基于弱监督学习的目标表观模型算法。在复杂场景下,由于剧烈光照、严重遮挡等因素的干扰,目标外观在新一帧中可能发生剧烈的本质的变化。为了使表观模型能较好地处理这些变化,我们利用简单高效的跟踪器对目标在新的一帧中进行粗定位,在这一位置采集弱标记的正样本加入到训练集中。在训练目标的表观模型时,通过流形正则项约束标记样本与弱标记样本之间的关系。训练集中包含了目标在新一帧外观信息,使学习得到的表观模型一定程度上能够反映目标在新的一帧中的外观变化。实验结果表明该算法性能优于其他的经典算法。 3. 在实验过程中发现单一特征无法处理目标在跟踪过程中的所有外观变化。为了增强特征的对目标的描述能力,利用增强核SVM融合不同的互补的特征学习目标的表观模型。同时为了增强表观模型对短时间内目标表观的剧烈变化的鲁棒性,我们对跟踪结果进行聚类,挑选出具有代表性的历史信息加入到训练集中学习目标的表观模型。通过对实验结果观测,该算法对目标短时间内的剧烈变化具有很强的鲁棒性。 4. 在分析现有基于稀疏表示的跟踪算法的缺点基础上,提出了基于鲁棒的多任务学习的目标跟踪算法。考虑到粒子点的相关性和差异性,在学习目标的稀疏系数表达时,将系数矩阵分解成两部分,分别利用联合稀疏与元素稀疏正则项对其进行惩罚。联合稀疏要求粒子点的重稀疏系数具有相同的结构,元素稀疏则对稀疏系数没有结构要求。矩阵分解得到的稀疏系数矩阵更精确,使得到的表观模型更鲁棒。实验表明该算法可以有效地处理跟踪过程中存在剧烈光照变化,严重遮挡和尺度变化等情况。
英文摘要Visual object tracking is an important problem in computer vision and has many applications including intelligent video surveillance monitoring, augmented reality and human computer interface. Although it has been investigated in the past decades, designing a robust tracker to cope with different objects under various situations is still a great challenging task. A very common difficulty is to resist the visual appearance changing frame by frame due to sudden illumination changes, background clutter, 3D rotation and partial occlusion. Such changes may make a tracker drift away from the target object. A typical tracking system consists of three components: 1) an appearance model, which evaluates the likelihood that the object of interest is at some particular location, 2) a motion model, which can relate the locations of the object over time, and 3) a search strategy for finding the most likely location in the current frame. Among them, the appearance model plays a crucial role. In this dissertation, we focus on the model-free tracking problem, i.e., no prior knowledge except for the object location is known at the beginning of tracking. The main components of our thesis are listed as follows: 1. A tracking algorithm based on Ranking SVM is proposed. During the tracking process, the tracker needs to label samples to update the appearance model. Unfortunately, it is hard to decide whether a sample should be labeled positive or negative. However, it is easy to decide the relative relation between patches. We tackle the tracking problem as ranking and employ ranking SVM to learn the relative order. Experiments show the superior performance of our tracker over several state-of-the-art tracking algorithms. 2. We propose a tracking algorithm based on weakly supervised learning. Most existing approaches utilize the object information contained in the current and previous frames to construct the object appearance model and locate the object with the model in frame t + 1. This method may work well if the object appearance just fluctuates in short time intervals. Nevertheless, suboptimal locations will be generated in frame t + 1 if the visual appearance changes substantially from the model under complex environment. To address this problem, a simple and effective tracker is employed to provide rough locations of the target object, i.e., some weakly labeled samples, in the next frame. Manifold regularization is used to model the relation between labeled and we...
语种中文
其他标识符201118014628027
源URL[http://ir.ia.ac.cn/handle/173211/6588]  
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
白延成. 复杂场景下视觉单目标跟踪算法研究[D]. 中国科学院自动化研究所. 中国科学院大学. 2014.

入库方式: OAI收割

来源:自动化研究所

浏览0
下载0
收藏0
其他版本

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。