2、万方数据Classified Index: TP391.4U.D.C: 004.9Dissertation for the Master’s Degree of EngineeringRESEARCHONKERNELREGRESSIONMODELSFORVISUALTRACKINGCandidate:Xiao MaSupervisor:Associate Prof. Zhenyu HeAcademicDegreeAppliedfor:Master’s Degree of EngineeringSpecialty:Computer TechnologyShenzhen Graduate Sch
3、oolAffiliation:DateofDefense:December, 2016Degree-Conferring-Institution:Harbin Institute of Technology万方数据哈尔滨工业大学工程硕士学位论文摘要由于目标在视频序列中通常经历复杂的表观形态变化,视频的目标跟踪一直是计算机视觉领域的一个复杂的问题。最近的一些研究表明,将目标跟踪建模为在频域实现的回归问题取得了很好的效果,在许多视频序列测试集上超过了过去的经典方法。与之相反,直接在时域实现回归模型通常被认为是一种性能有限的方法。因为追踪问题过程中训练样本有限,用传统回归模型在小样本上训练得到的回
5、在较长的帧间保留目标的多表观形态信息。本文实现了一种新的称为系数限制的排斥群LASSO的算法来形式化地生成此鉴别性模板。同时为了验证上述两种方法的有效性,本文在CVPR2013评价集上做了实验。实验囊括了30多种目标跟踪方法和70多个视频序列。实验结果表明,本文提出的方法取得了满意的跟踪效果并超过了许多经典的目标跟踪方法。关键词:目标跟踪;核方法;回归模型;群LASSO-I-万方数据哈尔滨工业大学工程硕士学位论文AbstractVisual tracking remains a challenging problem in computer vision due to the intrica
6、te variationof target appearances. Some progress made in recent years have revealed that correlation filters,which formulate the tracking process by creating a regressor in the frequency domain, haveachieved remarkable experimental results on a large amount of video tracking sequences. On thecontra
7、ry, building the regressor in the spatial domain directly has been considered as a limitedapproach since the number of training samples is restricted. And without sufficient training samples,the regressor w