基于压缩感知的视频追踪方法研究

发布时间：2018-03-07 05:04

本文选题：协同目标追踪　切入点：计算机视觉　出处：《上海师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：视频目标追踪作为计算机视觉领域一项核心技术,是后续情景感知、目标行为分析、视频搜索等各种高层视觉处理的基础。视频中存在目标尺度变化、光的明暗变化、遮挡、快速移动以及运动模糊等不确定的影响因素,所以寻求高效、鲁棒性的目标追踪模型一直以来都是具有挑战性的课题,目前的追踪算法大多是根据邻近帧的观测样本来更新表观模型。尽管有些算法已经有了不错的进展,但经常还是会遇到复杂背景下追踪目标丢失的问题,而且视频目标追踪过程中面临的自动初始化、前景和背景含有大量杂波、非刚性目标等问题也尚待解决,最近的一些研究工作都是针对以上问题开展的。本文在参考和研究大量前人的科研成果的基础上,针对视频的目标进行追踪,提出了两种新型的视频目标追踪检测算法:(1)基于稀疏协同的目标追踪算法。本文算法与其他同类算法相比更具有适应性,基于生成式模型计算候选样本和模板之间的相似度,最后将二者融合计算出候选样本的置信度。通过对模型的在线更新减少跟踪中的漂移问题,并能够自适应的对目标变化做出响应。(2)基于压缩感知理论的目标追踪算法。首先,在正负样本区域随机抽取不同尺度下图像的DRLBP特征;然后将高维的特征信息投射到低秩的压缩域,并根据此压缩域特征建立表观模型;最后,本文使用一个随机稀疏测量矩阵来压缩前景和背景目标,通过在压缩域中对追踪目标进行实时的在线更新追踪。追踪问题最终被转化成为一个使用朴素贝叶斯分类器的二元分类问题。本文提出的基于压缩感知的实时在线目标追踪方法,实验结果证明可以较快速、实时的完成目标的在线追踪,同时兼顾目标尺度变化、遮挡等。本文针对上述两种稀疏追踪算法,分别在不同的实验视频数据集上进行追踪验证。实验表明本文所提追踪算法的单帧追踪算法平均1ms甚至更低,实验表明在不同的追踪场景下本文算法均具有较强的实用性。
[Abstract]:As a core technology in the field of computer vision, video target tracking is the basis of all kinds of high-level visual processing, such as scene perception, target behavior analysis, video search and so on. Because of the uncertain factors such as fast movement and motion blur, it has been a challenging task to seek an efficient and robust target tracking model. Most of the current tracking algorithms update the apparent model based on the observation samples of adjacent frames. Although some algorithms have made good progress, they often encounter the problem of tracking target loss in complex background. In addition, the problems of automatic initialization, foreground and background contain a lot of clutter, non-rigid targets and so on are still to be solved in the process of video target tracking. Some of the recent research work has been carried out in response to the above problems. Based on the reference and study of a large number of previous scientific research achievements, this paper aims at tracking the video target. In this paper, two new video target tracking algorithms: 1) based on sparse collaboration are proposed. Compared with other similar algorithms, this algorithm is more adaptive, and computes the similarity between candidate samples and templates based on generative model. Finally, the confidence degree of candidate samples is calculated by combining the two methods. By updating the model online, the drift problem in tracking can be reduced, and the target tracking algorithm based on compressed perception theory can be adapted to respond to the change of target. The DRLBP features of images with different scales are randomly extracted in the positive and negative sample regions. Then the feature information of high dimension is projected to the compression domain of low rank, and an apparent model is established according to the features of the compressed domain. Finally, In this paper, a random sparse measurement matrix is used to compress the foreground and background targets. The tracking problem is transformed into a binary classification problem using naive Bayes classifier by updating the tracking target in the compressed domain in real time. Line target tracking method, The experimental results show that the online tracking of target can be completed quickly and in real time, taking into account the change of target scale, occlusion and so on. This paper aims at the above two sparse tracking algorithms. The experiments show that the single frame tracking algorithm of the proposed tracking algorithm is less than 1 Ms on average, and the experiment shows that the algorithm has strong practicability in different tracking scenarios.
【学位授予单位】：上海师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】