复杂场景下基于稀疏表示的视频目标跟踪方法研究

发布时间：2018-04-17 11:45

本文选题：计算机视觉 + 视频目标跟踪　；参考：《哈尔滨工业大学》2017年博士论文

【摘要】：视频目标跟踪是计算机视觉中的热点问题,在智能监控、汽车导航、高级人机交互等各个领域具有巨大发展潜力,然而受到目标姿势变化、形状改变、快速运动、噪声、遮挡、杂波背景等因素的影响,实现健壮目标跟踪仍面临极大挑战。基于稀疏表示的目标跟踪是近年来较为流行的一种目标跟踪方法,尤其在噪声、部分遮挡等场景下具有非常理想的跟踪效果。目前该方法面临效率和精度两大问题。在效率上,该方法对每个候选样本求解1最小化问题一次,因而计算量大,较为耗时;在精度上,1分解具有潜在的不稳定性,因而影响候选样本的稀疏表示精度,影响目标跟踪精度。本文针对以上两个问题,从以下四个方面进行研究,以提高基于稀疏表示的目标跟踪方法在各种复杂场景下的跟踪性能。首先,针对视频图像序列中被跟踪对象所包含的丰富多模态特征信息,先后提出两种不同融合多模态特征的目标跟踪方法。其一,对候选样本的各模态特征分别进行稀疏表示,计算重建误差,然后累加计算总重建误差,并最终确定目标。其二,针对前一方法忽视同一候选样本各模态特征间关联,从而带来跟踪偏差的问题,提出对同一候选样本的多模态特征进行联合稀疏表示,然后分别计算各模态特征的重建误差,并累加计算总重建误差,并据此计算观察概率,确定目标。其次,将候选样本与目标模板间的局部性关联融入基于稀疏表示的目标跟踪框架。首先根据每个候选样本与各目标模板之间的相似性对各目标模板进行加权,以反映它们之间的局部性关联。然后,对每个候选目标进行稀疏表示,使其更容易被与之相似的目标模板表示,从而减少跟踪偏差,提高跟踪精度。然后,提出一种基于多模态加权联合稀疏表示的目标跟踪方法。考虑到不仅候选样本与目标模板间存在着局部性关联、候选样本的多模态特征间存在关联、而且由于候选样本均采样于一较小区域范围内,候选样本间也存在着相似性关联,提出同时将候选样本与目标模板间的数据局部性、候选样本的多模态特征间关联、及候选样本间的相似性融入基于稀疏表示的目标跟踪框架,有效地提高了各种复杂场景下的目标跟踪精度。最后,提出基于核稀疏表示的方法以解决目标快速运动与模糊场景的目标跟踪问题。针对目标快速运动而产生模糊的情况,通过采用合适的核函数,将候选样本与目标模板非线性地映射到高维希尔伯特(Helbert)空间,然后进行稀疏表示,从而有效解决了基于稀疏表示方法不能解决候选样本与目标模板间所存在的非线性关联问题,提高了目标快速运动与模糊场景的目标跟踪精度。本课题充分考虑了视频目标跟踪过程中影响跟踪性能的各种因素,有效地提高了基于稀疏表示目标跟踪方法在各种复杂场景下的跟踪性能,为视频目标跟踪的进一步深入研究及实际应用奠定了基础。
[Abstract]:Video target tracking is a hot issue in computer vision. It has great potential in intelligent monitoring, vehicle navigation, advanced human-computer interaction and so on. However, it is affected by the changes of pose, shape, fast movement, noise, occlusion, etc.Due to clutter background and other factors, robust target tracking still faces great challenges.Target tracking based on sparse representation is a popular target tracking method in recent years, especially in noise, partial occlusion and other scenes.At present, the method is faced with two major problems: efficiency and precision.In terms of efficiency, the method solves 1 minimization problem once for each candidate sample, which results in a large amount of computation and time consuming, and has potential instability in accuracy, thus affecting the sparse representation accuracy of candidate samples.The target tracking accuracy is affected.In order to improve the tracking performance of the sparse representation based target tracking method in various complex scenarios, the following four aspects are studied in this paper.Firstly, aiming at the rich multi-modal feature information contained in the object being tracked in the video image sequence, two different target tracking methods are proposed to fuse the multi-modal features.Firstly, each modal feature of the candidate sample is represented sparsely, then the reconstruction error is calculated, then the total reconstruction error is calculated and the target is finally determined.Secondly, aiming at the problem that the former method neglects the correlation of each modal feature of the same candidate sample, which leads to the tracking deviation, a joint sparse representation of the multi-modal feature of the same candidate sample is proposed.Then the reconstruction error of each modal feature is calculated, and the total reconstruction error is accumulated. The observation probability is calculated accordingly and the target is determined.Secondly, the local association between candidate samples and target templates is incorporated into the sparse representation based target tracking framework.Firstly, each target template is weighted according to the similarity between each candidate sample and each target template to reflect the local association between them.Then, each candidate target is represented sparsely to make it easier to be represented by similar target template, so that the tracking deviation is reduced and the tracking accuracy is improved.Then, a target tracking method based on multimodal weighted joint sparse representation is proposed.Considering that there is not only a local correlation between candidate samples and target templates, but also a similar correlation between candidate samples because the candidate samples are sampled in a small region, and the multi-modal features of candidate samples are correlated.At the same time, the data locality between candidate samples and target templates, the association of multi-modal features of candidate samples, and the similarity between candidate samples are incorporated into the sparse representation based target tracking framework.The precision of target tracking in various complex scenes is improved effectively.Finally, a kernel sparse representation method is proposed to solve the problem of fast moving target and target tracking in fuzzy scene.In view of the ambiguity caused by the fast motion of the target, the candidate samples and the target template are mapped to the high dimensional Hilbert Helbert space by using the appropriate kernel function, and then the sparse representation is performed.Therefore, the nonlinear correlation between candidate samples and target templates can not be solved effectively based on sparse representation, and the target tracking accuracy of fast moving target and fuzzy scene is improved.In this paper, the factors that affect the tracking performance of video target tracking are fully considered, and the tracking performance of the sparse representation based target tracking method in various complex scenarios is improved effectively.It lays a foundation for further research and practical application of video target tracking.
【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】