欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 图像处理与计算机视觉 • 上一篇    下一篇

联合增强局部最大发生特征和k-KISSME 度量学习的行人再识别

  

  1. (1. 合肥工业大学计算机与信息学院，安徽 合肥 230009；
    2. 工业安全与应急技术安徽省重点实验室，安徽 合肥 230009)
  • 出版日期:2020-06-30 发布日期:2020-08-18
  • 基金资助:
    国家自然科学基金面上项目(61471154)；安徽省科技攻关强警项目(1704d0802181)；中央高校基本科研业务费专项资金资助项目(JZ2018YYPY0287)

Joint enhanced local maximal occurrence representation and k-KISSME metric learning for person re-identification

  1. (1. School of Computer and Information, Hefei University of Technology, Hefei Anhui 230009, China;
    2. Anhui Province Key Laboratory of Industry Safety and Emergency Technology, Hefei Anhui 230009, China)
  • Online:2020-06-30 Published:2020-08-18

PDF (PC)

116

摘要： 摘要：行人再识别是一种在监控视频中自动搜索行人的重要技术，该技术包含特征表示
和度量学习2 部分。有效的特征表示应对光线和视角变化具有鲁棒性，具有判别性的度量学习
能够提高行人图像的匹配精度。但是，现有的特征大多都是基于局部特征表示或者全局特征表
示，没有很好的集成行人外观的精细细节和整体外观信息且度量学习通常是在线性特征空间进
行，不能高效地利用特征空间中的非线性结构。针对该问题，设计了一种增强局部最大发生的
有效特征表示(eLOMO)方法，可以实现行人图像精细细节和整体外观信息的融合，满足人类视
觉识别机制；并提出一种被核化的KISSME 度量学习(k-KISSME)方法，其计算简单、高效，只
需要对2 个逆协方差矩阵进行估计。此外，为了处理光线和视角变化，应用了Retinex 变换和
尺度不变纹理描述符。实验表明该方法具有丰富和完整的行人特征表示能力，与现有主流方法
相比提高了行人再识别的识别率。

关键词: 行人再识别, 增强的局部最大发生特征, 核学习, 特征表示, 度量学习

Abstract: Person re-identification is an important technique for automatically searching for
pedestrians in surveillance videos. This technology consists of two key parts, feature representation
and metric learning. Effective feature representations should be robust to changes in illumination and
viewpoint, and the discriminative metric learning can improve the matching accuracy of person
images. However, most of the existing features were based on local or global feature representation
and failed to efficiently use the fine details and profile information of the appearance of pedestrians.
More importantly, metric learning was usually conducted in a linear feature space, and nonlinear
structures in the feature space couldn’t be efficiently utilized. To solve these problems, we first
designed an effective feature representation called enhanced local maximal occurrence representation
(eLOMO), which could realize the fusion of fine details and profile information of the appearance of the person image and satisfy the human visual recognition mechanism. Furthermore, we proposed a
kernelized KISSME metric learning (k-KISSME) method, simple and efficient, only requiring two
inverse covariance matrices to be estimated. In addition, to handle changes in light and viewing angle,
we applied Retinex transforms and scale-invariant texture descriptors. Experiments show that the
proposed method possesses the ability regarding abundant and integral person feature representation
and improves the recognition rate of person re-identification in comparison with the existing
mainstream methods.

Key words: person re-identification, enhanced local maximal occurrence feature, kernel-based learning, feature representation, metric learning