欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 图像与视频处理 • 上一篇    下一篇

基于迁移学习的行人再识别

  

  1. 1. 南通大学计算机科学与技术学院，江苏 南通 226019； 
    2. 南通先进通信技术研究院，江苏 南通 226019； 
    3. 南通大学交通学院，江苏 南通 226019
  • 出版日期:2018-10-31 发布日期:2018-11-16
  • 基金资助:
    江苏省教育厅自然科学基金项目(16KJB520037)；南通市前沿与关键技术创新项目(MS22015100)；江苏省社会安全图像与视频理解重点实验 室创新基金项目(30916014107)

Person Re-Identification Based on Transfer Learning

  1. 1. School of Computer Science and Technology, Nantong University, Nantong Jiangsu 226019, China; 
    2. Nantong Research Institute for Advanced Communication Technologies, Nantong Jiangsu 226019, China; 
    3. School of Transportation, Nantong University, Nantong Jiangsu 226019, China
  • Online:2018-10-31 Published:2018-11-16

PDF (PC)

171

摘要： 行人再识别指的是在无重叠的多摄像机监控视频中，匹配不同摄像机中的行人目 标。提出了一种基于迁移学习的行人再识别方法。在训练阶段，针对现有的基于深度卷积神经 网络的图像识别模型进行参数微调，将网络模型迁移学习至行人再识别模型。测试阶段，利用 学习好的网络模型提取行人图像的特征，再采用余弦距离来描述行人对之间的相似度。在 CUHK03、Market-1501 和 DukeMTMC-reID 3 个数据集上进行了深入的实验分析，实验结果表 明该方法取得了较高的累积匹配得分，特别是第 1 匹配率远远超过了非深度学习的方法，与其 他基于深度学习的行人再识别方法相比，准确率也有所提升。

关键词: 行人再识别, 深度卷积神经网络, 迁移学习

Abstract: The person re-identification refers to matching pedestrian images observed from different cameras in a non-overlapping multi-camera surveillance systems. In this article, a person re-identification method based on transfer learning is proposed. The deep convolutional neural network model pre-trained on ImageNet is adopted to fine-tune the parameters for person re-identification, and the Re-ID model is used to extract the features from person image. The simple cosine distance is applied to measure similarities between person pairs. The approach is evaluated by operating in-depth experiments in three benchmarks, including CUHK03, Market-1501 and DukeMTMC-reID, and the experimental results on the benchmarks show significant and consistent improvements over parallel methods.

Key words: person re-identification, deep convolutional neural network, Transfer learning