基于稀疏表示的视觉目标跟踪方法研究

发布时间：2018-12-16 13:46

【摘要】：视频序列中的目标跟踪问题是计算机视觉领域的研究热点,它融合了机器学习与模式识别等领域的研究成果,在视频监控、智能交通以及现代军事中得到了广泛的应用。针对目标跟踪问题,国内外的研究者们开展了大量的研究,并提出了许多行之有效的跟踪算法,但是如何在复杂多变的自然场景中对表观变化的目标进行鲁棒的跟踪仍然是一个极具挑战性的问题。常见的跟踪难点包括：场景光流的变化、目标尺度的变化、局部遮挡、目标的非刚性形变以及位姿的变化等。这些难点导致目标在视频中的表观呈现出非线性的变化,从而使跟踪问题变得更加复杂。本文针对上述难点问题,对目标跟踪问题进行了深入的研究,旨在提高目标跟踪算法的准确性和鲁棒性。本文的主要贡献如下：(1)提出了一种基于整体稀疏表示粒子预判的两步目标跟踪算法。针对大多数相关算法需要对全部采样粒子进行观测建模的问题,该算法在第一步采用整体模板的稀疏表示来建模粒子预判问题,并通过粒子预判去除偏离目标真实状态较远的粒子,能够有效减少样本的数目,提高算法效率。为了降低跟踪中发生漂移的可能性,该算法在第二步将目标的初始状态和当前状态同时作为观测参考,提高了跟踪结果的准确性。多个测试视频的跟踪实验结果表明了该算法的有效性。(2)提出了一种基于局部判别稀疏表示的目标跟踪算法。该算法针对现有相关算法对目标与背景区分能力不足的问题,增加了局部背景图像作为负样本来训练字典,兼顾了字典的表达能力和判别能力；针对现有局部稀疏模型缺少目标全局信息的问题,将目标建模为指定局部图像在字典空间上的稀疏编码直方图,这些局部图像构成了稀疏字典且具有一定的结构性,因此目标模型能够有效地联合目标的局部特征和全局结构特征。为了提高观测模型的准确性,设计了基于目标结构信息的相似性系数来衡量目标与样本间的相似性,并对目标模型进行主动的更新以适应跟踪中的表观变化。实验结果表明该算法能够处理大部分的跟踪难点问题,并且具有较高的跟踪精度。(3)提出了一种基于加权结构稀疏表示的目标跟踪算法。该算法采用结构稀疏表示对目标建模,能够充分利用目标局部图像间的结构信息,有效地避免了模型的退化；同时在结构稀疏字典中添加背景信息,能够增强模型对背景的判别能力；此外根据局部图像在表达目标时所起的作用分配重要性权重,并将目标建模为加权结构稀疏模型,极大地提高了模型的鲁棒性；为了减轻局部遮挡对跟踪的影响,在设计观测模型时增加了遮挡检测模块。在多个测试视频上的实验结果表明加权结构稀疏模型对目标表观变化具有很好的适应性,该算法在跟踪中表现出良好的鲁棒性和准确性。(4)提出了一种基于局部联合稀疏表示的多任务跟踪算法。该算法针对相关算法对样本结构信息利用不足的问题,将样本中局部图像的稀疏编码视为独立的任务,并在多任务学习框架下对所有样本的局部图像进行联合稀疏编码。这种联合稀疏编码能够最大程度地挖掘样本内部以及样本之间的结构关系,在增加模型表达能力的同时提高算法的工作效率。此外,还设计了联合相似性度量函数,从整体和局部两个方面衡量目标和样本之间的相似程度,提高了观测模型的可靠性。实验表明该算法在测试视频上具有更加准确的跟踪结果。综上,本文分别利用整体稀疏表示、局部判别稀疏表示、加权结构稀疏表示以及局部联合稀疏表示实现了视频序列的目标跟踪,并对这四种跟踪算法进行了实验分析,实验结果证明本论文渐进地提高了目标跟踪算法的准确性和鲁棒性。
[Abstract]:The target tracking problem in the video sequence is a hot topic in the field of computer vision. It combines the research results in the fields of machine learning and pattern recognition, and has been widely used in video surveillance, intelligent traffic and modern military. Aiming at the problem of target tracking, researchers at home and abroad have carried out a lot of research and put forward many effective tracking algorithms, but how to track the target of apparent change in the complex and changeable natural scene is still a challenging problem. The common tracking difficulties include the change of the scene light flow, the change of the target scale, the local occlusion, the non-rigid deformation of the target, and the change of the position. these difficulties lead to an apparent non-linear change in the target's appearance in the video, making the tracking problem more complex. The aim of this paper is to improve the accuracy and robustness of the target tracking algorithm. The main contributions of this paper are as follows: (1) A two-step target tracking algorithm based on the overall sparse representation of particle pre-judgment is presented. Aiming at the problem that most relevant algorithms need to observe and model all sampling particles, the algorithm uses the sparse representation of the whole template in the first step to model the particle pre-judgment problem, and the particles which deviate from the real state of the target are pre-determined by the particles, the number of samples can be effectively reduced, and the algorithm efficiency is improved. in order to reduce the possibility of drift in the tracking, the algorithm improves the accuracy of the tracking result by taking the initial state and the current state of the target as the observation reference in the second step. The experimental results of a plurality of test videos show the effectiveness of the algorithm. (2) A target tracking algorithm based on partial discrimination and sparse representation is proposed. aiming at the problem that the existing correlation algorithm is insufficient to distinguish the target and the background, the local background image is added as a negative sample to train the dictionary, the expression capability and the judgment capability of the dictionary are taken into account, the problem that the existing local sparse model lacks the target global information is solved, The target is modeled as a sparse coding histogram of a given local image in a dictionary space, which form a sparse dictionary and have a certain structure, so that the target model can effectively combine the local features and the global structural features of the target. In order to improve the accuracy of the observation model, a similarity coefficient based on the target structure information is designed to measure the similarity between the target and the sample, and the target model is actively updated to meet the apparent change in the tracking. The experimental results show that the algorithm can deal with most of the tracking difficulties and has higher tracking precision. (3) A target tracking algorithm based on the sparse representation of the weighted structure is proposed. The method adopts the structure sparse representation to model the target, can fully utilize the structure information between the target local images, effectively avoids the degradation of the model, and simultaneously adds the background information in the structure sparse dictionary, so that the discrimination capability of the model on the background can be enhanced; in addition, the importance weight is distributed according to the action of the local image when the target is expressed, and the target is modeled as a weighted structure sparse model, so that the robustness of the model is greatly improved; and in order to reduce the influence of the local occlusion on the tracking, the occlusion detection module is added when the observation model is designed. The experimental results show that the weighted-structure sparse model has good adaptability to the apparent change of the target, and the algorithm shows good robustness and accuracy in the tracking. (4) A multi-task tracking algorithm based on local joint sparse representation is proposed. Aiming at the problem that the correlation algorithm has insufficient use of the sample structure information, the sparse coding of the local image in the sample is regarded as an independent task, and the partial images of all the samples are combined and sparse coded under the multi-task learning framework. The combined sparse coding can maximize the structural relationship between the samples and the samples, and improve the working efficiency of the algorithm while increasing the expression ability of the model. In addition, a joint similarity measure function is designed to measure the similarity between the target and the sample from both the whole and the local aspects, and the reliability of the observation model is improved. The experiment shows that the algorithm has more accurate tracking results on the test video. In this paper, the target tracking of the video sequence is realized by the sparse representation of the whole, the sparse representation of the local discrimination, the sparse representation of the weighted structure and the local joint sparse representation, and the four tracking algorithms are analyzed. The experimental results show that this paper improves the accuracy and robustness of the target tracking algorithm.
【学位授予单位】：大连理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】