基于稀疏表示的在线字典学习模型目标跟踪算法研究

发布时间：2018-05-25 20:37

本文选题：目标跟踪 + 外观模型　；参考：《江南大学》2017年硕士论文

【摘要】：目标跟踪是计算机视觉领域的核心内容,在视频监控、人机交互、交通检测等领域得到了重要的应用。尽管已经展开广泛的研究,并提出了许多优秀的算法,但是建立一个稳定、高效的跟踪系统仍然是一个非常具有挑战性的任务,目前目标跟踪面临的困难主要有姿势变化、光照变化、遮挡、突然运动等。本文中,我们通过建立几个鲁棒的外观模型来解决这些困难。1、为了提高目标跟踪速度,提出一种基于兴趣点的局部外观模型。包括两个部分:基于兴趣点的目标表示和基于稀疏表示的鲁棒互相匹配原则。基于兴趣点的目标表示是以兴趣点为中心,提取一定尺寸的图像块组成目标字典。在随后各帧中,以各个候选兴趣点为中心利用粒子滤波算法提取不同方向和尺度的候选目标块。基于稀疏表示的互相匹配原则利用稀疏表示方法计算目标字典和候选目标块之间的对应关系,并筛选候选目标块。最后利用所有选定的候选块平均偏移量来确定目标位置。另外,提出一种在线字典学习方法更新目标模板,可以有效的将过去目标信息和当前跟踪到的目标信息融合在一起,融合的过程中能够自动的降低外界干扰信息添加到目标模板中的概率,从而降低目标模板漂移的概率。2、为了增强目标跟踪对干扰的抵抗力,首先提出局部自适应加权算法来增加受干扰影响的目标区域与未受干扰影响的目标区域之间的区分度。然后提出结构稀疏表示算法,利用包含丰富目标特征和背景特征的候选目标构建过完备字典来重构目标模板。再利用系数矩阵内部的结构性统计每个候选目标与目标模板之间的相似度,可以提高跟踪的精度。最后提出在线判别双字典学习算法,这种字典学习方法在融合目标过去信息和当前信息的过程中通过判别函数能够自动检测并排斥背景信息,使更新后的目标模板能够精确的反映目标变化。本文深入研究了基于粒子滤波框架的稀疏表示目标跟踪算法,提高了跟踪算法在具有各种干扰因素影响的复杂环境中的实时性、稳定性和准确性。
[Abstract]:Target tracking is the core of the field of computer vision. It has been widely used in video surveillance, human-computer interaction, traffic detection and other fields. Although extensive research has been carried out and many excellent algorithms have been proposed, a stable and efficient tracking system is still a very challenging task at present. The main difficulties faced by tracking are postural change, illumination change, occlusion, and sudden motion. In this paper, we solve these difficult.1 by building several robust appearance models. In order to improve the target tracking speed, we propose a local appearance model based on interest points. It includes two parts: object representation based on interest points and based on dilute. The target representation based on interest points is centered on interest points and extracts a certain size of image blocks to form a target dictionary. In the subsequent frames, the candidate target blocks with different directions and scales are extracted with particle filter algorithm at the center of each candidate interest point. The sparse representation method is used to calculate the correspondence between the target dictionary and the candidate target block, and to select the candidate target blocks. Finally, the target location is determined by using the average offset of all selected candidates. In addition, an online dictionary learning method is proposed to update the target template, which can effectively track the past target information and track the target information to the current target. In the process of fusion, the probability of the external interference information added to the target template can be automatically reduced and the probability.2 of the target template drift can be reduced. In order to enhance the resistance of the target tracking to the interference, the local adaptive weighting algorithm is first proposed to increase the target area and unaffected interference affected by the interference. The structural sparse representation algorithm is proposed, and an overcomplete dictionary is constructed to reconstruct the target template with a candidate target containing rich target features and background features, and the similarity between each candidate target and the target template can be improved by using the internal structure of the coefficient matrix to improve the heel. In the end, the online discriminant double dictionary learning algorithm is proposed. This dictionary learning method can automatically detect and exclude background information through the discriminant function in the process of merging the target information and the current information, so that the updated target template can accurately reflect the change of the target. This paper deeply studies the particle filter frame based on the particle filter frame. The sparse representation of target tracking algorithm improves the real-time performance, stability and accuracy of tracking algorithm in complex environment with various interference factors.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】