【作者】 李鹏程；

【导师】 李映；

【作者基本信息】 西北工业大学 ， 计算机应用技术， 2015， 硕士

【摘要】 视觉跟踪作为计算机视觉领域中的一个基础性问题,是一项融合了数字信号处理、模式识别、机器学习、人工智能、数据挖掘、多媒体检索等许多相关领域先进成果的高级课题。在民用和军用的许多领域有着广阔的应用前景。近年来,研究人员已经提出了许多经典的视觉跟踪算法,但是设计一个能够兼顾目标的各种视觉外观变化、遮挡与噪声、尺度变化和复杂的运动背景的跟踪算法,仍然是一个具有挑战性的难题。本文回顾了贝叶斯框架下的粒子滤波算法,利用稀疏表示理论,基于主成分分析的子空间模型,多任务学习,朴素贝叶斯分类器等方法,提出了三种目标跟踪算法。1)在原始的高维度Haar-like特征基础上,利用稀疏表示的特征选择作用,对其进行降维处理,选取那些对于目标和背景区分有显著效果的特征。选取正负样本降维后的特征,训练朴素贝叶斯分类器。最后,利用该分类器进行目标跟踪。对比实验表明,该方法在具有挑战性的视频序列上取得了较好的跟踪效果。2)在传统的基于稀疏表示的目标跟踪算法框架下,引入基于PCA分析的子空间外观模型更好的表示目标特征,且利用增量PCA算法进行外观模型的在线更新,较好地处理目标的外观变化。同时,将跟踪算法中每个待选目标的稀疏求解过程视为一个独立的任务,借助多任务学习的方法,挖掘各个任务间的关系,联合求解出稀疏表示的系数。在多组视频序列上的对比实验表明,该方法具有较好的跟踪效果。3)将稀疏表示方法应用到目标的外观模型表示中,构建了一个基于稀疏表示的局部外观模型。此模型充分利用了稀疏表示的系数来刻画目标,能够在目标与背景之间做出较好的区分。同时,增量子空间学习和稀疏表示共同应用于更新目标模型,能够较好地处理漂移和部分遮挡的问题。借助传统的基于稀疏表示的目标跟踪算法框架,该方法在多组视频序列上均取得了较好的跟踪效果。更多还原

【Abstract】 Visual tracking is a fundamental problem in the field of computer vision, which includes many other advanced technologies such as digital signal processing, pattern recognition, machine learning, data mining, multimedia search and so on. There are broad applications about tracking in both military and civil fields. In recent years, researchers have provided a lot of tracking methods. But how to design a good tracking algorithm which can handle with the appearance changes, occlusion, noise, and scale change is still a challenging problem.In this paper, we review the particle filer algorithm in the Bayes framework and adopt the subspace model based on principal component analysis, sparse representation, multi-task learning, na?ve Bayes classifier to provide three visual tracking methods.1) We extract the Haar-like features from the object and then use the sparse representation to select those features which are good at distinguishing the object and background. In the following step, we train the na?ve Bayes classifier based on positive and negative samples. In the end, we adopt this classifier to track the object and acquire better performance.2) We introduce the subspace model based on PCA into the traditional L1 tracking framework. Also, the incremental PCA method is applied to update the appearance model. So, we can better represent the object. In addition, we view the sparse representation process of each candidate object as one task and use the technology of multi-task learning to mine the relationship among the tasks. Therefore, we improve the tracking accuracy.3) We apply the sparse representation into appearance model and establish a local appearance model based on sparse representation. This model takes full advantages of the coefficients of sparse representation to capture the object and can better distinguish between the object and background. Meantime, the incremental PCA and sparse representation are both used in updating the appearance model and can deal with the occlusion and drift problem. This model with the L1 framework can do a good job.更多还原