【作者】 张晶；

【导师】 赵旭；

【作者基本信息】 上海交通大学 ， 控制科学与工程， 2017， 硕士

【摘要】 人体目标再识别是计算机视觉和智能监控领域的重点研究课题之一,其任务是匹配无重叠视域的摄像机在不同时间、地点拍摄的人体目标。受光照条件、背景、遮挡、视角和姿态等因素影响,不同摄像机下的同一目标表观差异较大。目前研究主要集中在特征表示和度量学习两方面:一方面,研究者们致力于提出区分度高、鲁棒性强的特征描述子;另一方面,很多度量学习方法在人体目标再识别问题上了取得了较好的效果。但对于多样化的数据集,单一的全局度量很难适应差异化的特征。有研究者提出了局部度量学习思想,但由于再识别问题本身的研究挑战,现有的局部度量学习方法无法直接应用于该问题。本文以提高在多样化数据集上的再识别准确率为出发点,利用局部度量学习思想,结合现有的全局度量学习算法,提出一系列多度量学习方法,有效提高了再识别的性能,具有广泛的应用价值。具体包括以下几项研究工作:1)提出基于高斯混合模型和半正定约束的多度量学习方法。将度量矩阵约束为半正定矩阵,以最小化训练集的对数逻辑损失函数为目标。使用高斯混合模型拟合训练样本分布,并利用样本在各个高斯成分下的后验概率,改进目标损失函数中不同样本的损失权重,每个高斯成分对应不同的优化目标,对多个目标函数分别使用加速近端梯度法进行迭代优化,学习多个度量矩阵,以此计算多个相似性距离,并将多个距离加权结合。2)有效地将局部度量学习方法应用到再识别问题,提出以聚类间存在样本重叠的子集划分方式进行局部训练集划分,有效克服了局部度量学习方法在人体目标再识别问题中的应用困难。3)利用局部度量学习思想,结合近几年提出的交叉视角二次判别分析(Cross-view Quadratic Discriminant Analysis,XQDA)和加速近端梯度度量学习(Metric Learning by Accelerated Proximal Gradient,MLAPG)等全局度量学习方法,提出一种整合全局-局部度量学习框架。利用高斯混合模型对训练样本进行聚类,并使用2)中提出策略划分局部训练子集,在每个局部子集内分别进行局部度量学习;同时在整个训练集上进行全局度量学习。对于测试样本,根据样本在高斯混合模型各成分下的后验概率,将局部和全局度量矩阵加权结合,做为衡量样本相似性的依据。在VIPeR、PRID 450S和Market-1501等数据集上的实验结果验证了本文提出的基于高斯混合模型的多度量学习、局部度量学习和整合全局-局部度量学习方法的有效性。在包含丰富的背景和人物衣着变化的VIPeR数据集上的匹配准确率达到42.0%以上,相比于XQDA和MLAPG等全局方法,提高2.0%以上,在其它数据集上的性能也有不同程度的提高,本文提出的方法的性能普遍优于或接近现研究阶段的主流方法。另外,在使用不同的特征表示条件下,提出的整合全局-局部度量学习框架均可对全局度量学习方法做出改进,该方法具有广泛的应用价值。更多还原

【Abstract】 Person re-identification is an important research subject of computer vision and intelligent surveillance,the task of which is to match snapshots of people from non-overlapping camera views at different time and places.Intra-class images from different cameras show different appearances due to variations of illumination,backgrounds,occlusions,viewpoints and poses.Feature representation and metric learning are two major research directions of person re-identification.On one hand,some researches focus on feature descriptors which are discriminative for different classes and robust against intra-class variations.On the other hand,many metric learning algorithms have achieved good performance on person re-identification.However,comparing all of the samples with a single global metric is inappropriate to handle heterogeneous dataset.Some researchers propose local metric learning.But these methods cannot be directly used for person re-identification due to some research challenges.In order to improve the matching performance on heterogeneous data,a series of novel multi-metric learning approaches which combine the idea of local metric learning and some existing global metric learning algorithms are proposed.The proposed methods improve the matching accuracy and can be widely used on person re-identification.The specific research works are as follows:1)A multi-metric learning approach based on Gaussian mixture model(GMM)and positive semi-definite(PSD)constraint is proposed.The approach is aimed at minimizing a log logistic loss function of the training set.GMM is used to fit the distribution of training samples.The posterior probabilities of each Gaussian component are used to improve the weights of different sample pairs’ loss in the entire loss function.Each component is corresponding to a different optimizing objective function with PSD constraint for the metric matrix.Accelerated proximal gradient algorithm is used to solve the optimizing problems.And multiple metrics are obtained,which are used to compute multiple similaritiy distances to be combined.2)The idea of local metric learning is used on the person re-identification problems by partitioning the entire training set into multiple local subsets with overlapping samples.It successfully overcomes the research challenges and difficulties of using local metric learning on person re-identification issue.3)Local metric learning is combined with some recently proposed global metric learning approaches such as cross-view quadratic discriminant analysis(XQDA)and a PSD constrained asymmetric metric learning approach termed as MLAPG.In the training stage,all of the samples are partitioned into several clusters by GMM softly.Also,the dividing strategy proposed in 2)is used.Local metrics are learned on each subset respectively by existing metric learning methods.Meanwhile,a global metric is also learned on the entire training set.In the testing stage,for each pair of samples,the local metrics weighted by their posterior probabilities aligned to different GMM components and the global metric weighted by a cross-validated parameter are integrated into the final metric for similarity evaluation.The experimental results on three challenging datasets of person re-identification(VIPeR,PRID 450 S and Market-1501)show the effectiveness of the proposed multi-metric learning,local metric learning and integrated global-local metric learning approaches.Especially,on VIPeR dataset with large variations of backgrounds and clothes,the proposed approaches perform much better.The matching accuracy achieves at more than 42.0%,over 2.0% increase comparing to the exsiting global metric learning methods.The proposed methods perform better than or equal to the state-of-the-art methods.In addition,the approaches could be generalized to different scenarios,as they can improve the performance regardless of using what feature descriptors.更多还原