整合全局——局部度量学习的人体目标再识别

发布时间：2018-06-16 19:22

本文选题：人体目标再识别 + 度量学习　；参考：《中国图象图形学报》2017年04期

【摘要】：目的人体目标再识别的任务是匹配不同摄像机在不同时间、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,不同摄相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。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。对此,有研究者提出了局部度量学习,但这些方法通常需要求解复杂的凸优化问题,计算繁琐。方法利用局部度量学习思想,结合近几年提出的XQDA(cross-view quadratic discriminant analysis)和MLAPG(metric learning by accelerated proximal gradient)等全局度量学习方法,提出了一种整合全局和局部度量学习框架。利用高斯混合模型对训练样本进行聚类,在每个聚类内分别进行局部度量学习;同时在全部训练样本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个成分下的后验概率将局部和全局度量矩阵加权结合,作为衡量相似性的依据。特别地,对于MLAPG算法,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,进一步提高该方法的性能。结果在VIPeR、PRID 450S和QMUL GRID数据集上的实验结果验证了提出的整合全局—局部度量学习方法的有效性。相比于XQDA和MLAPG等全局方法,在VIPeR数据集上的匹配准确率提高2.0%左右,在其他数据集上的性能也有不同程度的提高。另外,利用不同的特征表示对提出的方法进行实验验证,相比于全局方法,匹配准确率提高1.3%3.4%左右。结论有效地整合了全局和局部度量学习方法,既能对多种全局度量学习算法的性能做出改进,又能避免局部度量学习算法复杂的计算过程。实验结果表明,对于使用不同的特征表示,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。
[Abstract]:Objective the task of human target recognition is to match the human target taken by different camera at different time and place. Under the influence of illumination conditions, background, occlusion, angle of view and posture, the same target appearance is different under different cameras. At present, the research focuses on two aspects: feature representation and metric learning. Many metric learning methods have achieved good results in the problem of human body target re-recognition, but for a variety of data sets, a single global metric is difficult to adapt to the characteristics of differentiation. Some researchers have proposed local metric learning, but these methods usually need to solve complex convex optimization problems. Methods using the idea of local metric learning and combining the global metric learning methods such as XQDA-cross-view quadratic discriminant analysis) and MLAPGmetric learning by accelerated proximal gradient) proposed in recent years, a framework for integrating global and local metric learning is proposed. The Gao Si hybrid model is used to cluster the training samples, and the local metric learning is carried out in each cluster, and the global metric learning is carried out on all the training sample sets. For the test samples, the local and global measurement matrices are weighted according to the posterior probability of the samples under each component of the Gao Si mixed model as the basis for measuring similarity. In particular, for the MLAPG algorithm, the loss weights of different samples in the target loss function are improved by using the posteriori probability of the samples under each Gao Si component, and the performance of the method is further improved. Results the experimental results on the VIPeRNPRID450S and QMUL-GRID datasets verify the effectiveness of the proposed integrated global-local metric learning method. Compared with the global methods such as XQDA and MLAPG, the matching accuracy on VIPeR dataset is improved by about 2.0%, and the performance on other datasets is improved to some extent. In addition, the experimental results show that the matching accuracy of the proposed method is about 3.4% higher than that of the global method. Conclusion effective integration of global and local metric learning methods can not only improve the performance of various global metric learning algorithms, but also avoid the complex computing process of local metric learning algorithms. The experimental results show that the proposed integrated global-local metric learning framework can improve the global metric learning method.
【作者单位】：上海交通大学自动化系系统控制与信息处理教育部重点实验室;
【基金】：国家自然科学基金项目(61273285,61673269,61375019)~~
【分类号】：TP181

【相似文献】